def get_vibe_results(pred_cam, pred_verts, pred_pose, pred_betas,
pred_joints3d, joints2d_lis, bbox_lis, frame_lis,
orig_dim, person_id):
pred_cam = torch.cat(pred_cam, dim=0).cpu().numpy()
pred_verts = torch.cat(pred_verts, dim=0).cpu().numpy()
pred_pose = torch.cat(pred_pose, dim=0).cpu().numpy()
pred_betas = torch.cat(pred_betas, dim=0).cpu().numpy()
pred_joints3d = torch.cat(pred_joints3d, dim=0).cpu().numpy()
orig_height, orig_width = orig_dim
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bbox_lis,
img_width=orig_width,
img_height=orig_height)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d_lis,
'bboxes': bbox_lis,
'frame_ids': frame_lis,
}
vibe_results = {}
vibe_results[person_id] = output_dict
return vibe_results
def main(args):
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
video_file = args.vid_file
# ========= [Optional] download the youtube video ========= #
if video_file.startswith('https://www.youtube.com'):
print(f'Donwloading YouTube video \"{video_file}\"')
video_file = download_youtube_clip(video_file, '/tmp')
if video_file is None:
exit('Youtube url is not valid!')
print(f'YouTube Video has been downloaded to {video_file}...')
if not os.path.isfile(video_file):
exit(f'Input video \"{video_file}\" does not exist!')
output_path = os.path.join(
args.output_folder,
os.path.basename(video_file).replace('.mp4', ''))
os.makedirs(output_path, exist_ok=True)
image_folder, num_frames, img_shape = video_to_images(video_file,
return_info=True)
print(f'Input video number of frames {num_frames}')
orig_height, orig_width = img_shape[:2]
total_time = time.time()
# ========= Run tracking ========= #
bbox_scale = 1.1
if args.tracking_method == 'pose':
if not os.path.isabs(video_file):
video_file = os.path.join(os.getcwd(), video_file)
tracking_results = run_posetracker(video_file,
staf_folder=args.staf_dir,
display=args.display)
else:
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = joints2d = None
if args.tracking_method == 'bbox':
bboxes = tracking_results[person_id]['bbox']
elif args.tracking_method == 'pose':
joints2d = tracking_results[person_id]['joints2d']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
dataloader = DataLoader(dataset,
batch_size=args.vibe_batch_size,
num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= [Optional] run Temporal SMPLify to refine the results ========= #
if args.run_smplify and args.tracking_method == 'pose':
norm_joints2d = np.concatenate(norm_joints2d, axis=0)
norm_joints2d = convert_kps(norm_joints2d, src='staf', dst='spin')
norm_joints2d = torch.from_numpy(norm_joints2d).float().to(device)
# Run Temporal SMPLify
update, new_opt_vertices, new_opt_cam, new_opt_pose, new_opt_betas, \
new_opt_joints3d, new_opt_joint_loss, opt_joint_loss = smplify_runner(
pred_rotmat=pred_pose,
pred_betas=pred_betas,
pred_cam=pred_cam,
j2d=norm_joints2d,
device=device,
batch_size=norm_joints2d.shape[0],
pose2aa=False,
)
# update the parameters after refinement
print(
f'Update ratio after Temporal SMPLify: {update.sum()} / {norm_joints2d.shape[0]}'
)
pred_verts = pred_verts.cpu()
pred_cam = pred_cam.cpu()
pred_pose = pred_pose.cpu()
pred_betas = pred_betas.cpu()
pred_joints3d = pred_joints3d.cpu()
pred_verts[update] = new_opt_vertices[update]
pred_cam[update] = new_opt_cam[update]
pred_pose[update] = new_opt_pose[update]
pred_betas[update] = new_opt_betas[update]
pred_joints3d[update] = new_opt_joints3d[update]
elif args.run_smplify and args.tracking_method == 'bbox':
print(
'[WARNING] You need to enable pose tracking to run Temporal SMPLify algorithm!'
)
print('[WARNING] Continuing without running Temporal SMPLify!..')
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
# Runs 1 Euro Filter to smooth out the results
if args.smooth:
min_cutoff = args.smooth_min_cutoff # 0.004
beta = args.smooth_beta # 1.5
print(
f'Running smoothing on person {person_id}, min_cutoff: {min_cutoff}, beta: {beta}'
)
pred_verts, pred_pose, pred_joints3d = smooth_pose(
pred_pose, pred_betas, min_cutoff=min_cutoff, beta=beta)
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
del model
end = time.time()
fps = num_frames / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
total_time = time.time() - total_time
print(
f'Total time spent: {total_time:.2f} seconds (including model loading time).'
)
print(
f'Total FPS (including model loading time): {num_frames / total_time:.2f}.'
)
print(
f'Saving output results to \"{os.path.join(output_path, "vibe_output.pkl")}\".'
)
joblib.dump(vibe_results, os.path.join(output_path, "vibe_output.pkl"))
if not args.no_render:
# ========= Render results as a single video ========= #
renderer = Renderer(resolution=(orig_width, orig_height),
orig_img=True,
wireframe=args.wireframe)
output_img_folder = f'{image_folder}_output'
os.makedirs(output_img_folder, exist_ok=True)
if args.joints3dview:
output_img_raw_folder = f'{image_folder}_raw_output'
os.makedirs(output_img_raw_folder, exist_ok=True)
output_img_joints3d_folder = f'{image_folder}_joints3d_output'
os.makedirs(output_img_joints3d_folder, exist_ok=True)
output_img_mesh_folder = f'{image_folder}_mesh_output'
os.makedirs(output_img_mesh_folder, exist_ok=True)
output_img_meshside_folder = f'{image_folder}_meshside_output'
os.makedirs(output_img_meshside_folder, exist_ok=True)
output_img_all_folder = f'{image_folder}_all_output'
os.makedirs(output_img_all_folder, exist_ok=True)
print(f'Rendering output video, writing frames to {output_img_folder}')
# prepare results for rendering
frame_results = prepare_rendering_results(vibe_results, num_frames)
mesh_color = {
k: colorsys.hsv_to_rgb(np.random.rand(), 0.5, 1.0)
for k in vibe_results.keys()
}
image_file_names = sorted([
os.path.join(image_folder, x) for x in os.listdir(image_folder)
if x.endswith('.png') or x.endswith('.jpg')
])
length_image_files = len(image_file_names)
#length_image_files = 100
for frame_idx in tqdm(range(length_image_files)):
img_fname = image_file_names[frame_idx]
img = cv2.imread(img_fname)
if args.sideview:
side_img = np.zeros_like(img)
if args.joints3dview:
img_raw = img.copy()
img_joints3d = np.zeros_like(img)
joints3d_list = []
for person_id, person_data in frame_results[frame_idx].items():
frame_verts = person_data['verts']
frame_cam = person_data['cam']
joints3d = person_data['joints3d']
#print('frame_verts.shape = {}\nframe_cam.shape ={}\njoints3d.shape = {}'.format(
# frame_verts.shape, frame_cam.shape, joints3d.shape))
mc = mesh_color[person_id]
if args.joints3dview:
joints3d_list.append(joints3d)
# img_joints3d = render_joints3d(joints3d, img_raw.shape)
mesh_filename = None
if args.save_obj:
mesh_folder = os.path.join(output_path, 'meshes',
f'{person_id:04d}')
os.makedirs(mesh_folder, exist_ok=True)
mesh_filename = os.path.join(mesh_folder,
f'{frame_idx:06d}.obj')
img = renderer.render(
img,
frame_verts,
cam=frame_cam,
color=mc,
mesh_filename=mesh_filename,
)
if args.sideview:
side_img = renderer.render(
side_img,
frame_verts,
cam=frame_cam,
color=mc,
angle=270,
axis=[0, 1, 0],
)
if args.sideview:
img_mesh = img.copy()
img = np.concatenate([img, side_img], axis=1)
cv2.imwrite(
os.path.join(output_img_folder, f'{frame_idx:06d}.png'), img)
if args.joints3dview:
#img_joints3d = np.zeros_like(img_raw)
if len(joints3d_list) == 0:
img_joints3d = np.zeros_like(img_raw)
else:
joints3d = np.concatenate(joints3d_list)
img_joints3d = render_joints3d(joints3d, img_raw.shape)
if args.joints3dview:
img_up = np.concatenate([img_raw, img_joints3d], axis=1)
img_down = np.concatenate([img_mesh, side_img], axis=1)
img_all = np.concatenate([img_up, img_down], axis=0)
cv2.imwrite(
os.path.join(output_img_raw_folder,
f'{frame_idx:06d}.png'), img_raw)
cv2.imwrite(
os.path.join(output_img_joints3d_folder,
f'{frame_idx:06d}.png'), img_joints3d)
cv2.imwrite(
os.path.join(output_img_mesh_folder,
f'{frame_idx:06d}.png'), img_mesh)
cv2.imwrite(
os.path.join(output_img_meshside_folder,
f'{frame_idx:06d}.png'), side_img)
cv2.imwrite(
os.path.join(output_img_all_folder,
f'{frame_idx:06d}.png'), img_all)
if args.display:
cv2.imshow('Video', img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if args.display:
cv2.destroyAllWindows()
# ========= Save rendered video ========= #
vid_name = os.path.basename(video_file)
save_name = f'{vid_name.replace(".mp4", "")}_vibe_result.mp4'
save_name = os.path.join(output_path, save_name)
print(f'Saving result video to {save_name}')
images_to_video(img_folder=output_img_folder,
output_vid_file=save_name)
shutil.rmtree(output_img_folder)
if args.joints3dview:
'''
save_name_raw = f'{vid_name.replace(".mp4", "")}_raw.mp4'
save_name_raw = os.path.join(output_path, save_name_raw)
images_to_video(img_folder=output_img_raw_folder, output_vid_file=save_name_raw)
shutil.rmtree(output_img_raw_folder)
save_name_joints3d = f'{vid_name.replace(".mp4", "")}_joints3d.mp4'
save_name_joints3d = os.path.join(output_path, save_name_joints3d)
images_to_video(img_folder=output_img_joints3d_folder, output_vid_file=save_name_joints3d)
shutil.rmtree(output_img_joints3d_folder)
save_name_mesh = f'{vid_name.replace(".mp4", "")}_mesh.mp4'
save_name_mesh = os.path.join(output_path, save_name_mesh)
images_to_video(img_folder=output_img_mesh_folder, output_vid_file=save_name_mesh)
shutil.rmtree(output_img_mesh_folder)
save_name_meshside = f'{vid_name.replace(".mp4", "")}_meshside.mp4'
save_name_meshside = os.path.join(output_path, save_name_meshside)
images_to_video(img_folder=output_img_meshside_folder, output_vid_file=save_name_meshside)
shutil.rmtree(output_img_meshside_folder)
'''
save_name_all = f'{vid_name.replace(".mp4", "")}_all.mp4'
save_name_all = os.path.join(output_path, save_name_all)
images_to_video(img_folder=output_img_all_folder,
output_vid_file=save_name_all)
shutil.rmtree(output_img_all_folder)
shutil.rmtree(image_folder)
print('================= END =================')
def run_vibe(video_file, args):
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
# Make output dirs
output_path = os.path.join(
args.output_folder, os.path.basename(video_file).replace('.mp4', ''))
os.makedirs(output_path, exist_ok=True)
# Convert video to images
image_folder, num_frames, img_shape = video_to_images(
video_file, return_info=True)
print(f'Input video number of frames {num_frames}')
orig_height, orig_width = img_shape[:2]
total_time = time.time()
# ========= Run tracking ========= #
if not os.path.isabs(video_file):
video_file = os.path.join(os.getcwd(), video_file)
tracking_results = run_posetracker(
video_file, staf_folder=args.staf_dir, display=args.display, smoothen=args.smoothen, smoothen_method=args.smoothen_method)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
joints2d = tracking_results[person_id]['joints2d']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=None,
joints2d=joints2d
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
dataloader = DataLoader(
dataset, batch_size=args.vibe_batch_size, num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [
], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= [Optional] run Temporal SMPLify to refine the results ========= #
norm_joints2d = np.concatenate(norm_joints2d, axis=0)
norm_joints2d = convert_kps(norm_joints2d, src='staf', dst='spin')
norm_joints2d = torch.from_numpy(norm_joints2d).float().to(device)
# Run Temporal SMPLify
update, new_opt_vertices, new_opt_cam, new_opt_pose, new_opt_betas, \
new_opt_joints3d, new_opt_joint_loss, opt_joint_loss = smplify_runner(
pred_rotmat=pred_pose,
pred_betas=pred_betas,
pred_cam=pred_cam,
j2d=norm_joints2d,
device=device,
batch_size=norm_joints2d.shape[0],
pose2aa=False,
)
# update the parameters after refinement
print(
f'Update ratio after Temporal SMPLify: {update.sum()} / {norm_joints2d.shape[0]}')
pred_verts = pred_verts.cpu()
pred_cam = pred_cam.cpu()
pred_pose = pred_pose.cpu()
pred_betas = pred_betas.cpu()
pred_joints3d = pred_joints3d.cpu()
pred_verts[update] = new_opt_vertices[update]
pred_cam[update] = new_opt_cam[update]
pred_pose[update] = new_opt_pose[update]
pred_betas[update] = new_opt_betas[update]
pred_joints3d[update] = new_opt_joints3d[update]
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
orig_cam = convert_crop_cam_to_orig_img(
cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height
)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
del model
end = time.time()
fps = num_frames / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
total_time = time.time() - total_time
print(
f'Total time spent: {total_time:.2f} seconds (including model loading time).')
print(
f'Total FPS (including model loading time): {num_frames / total_time:.2f}.')
print(
f'Saving output results to \"{os.path.join(output_path, "vibe_output.pkl")}\".')
# joblib.dump(vibe_results, os.path.join(output_path, "vibe_output.pkl"))
for person in vibe_results.keys():
dump_path = os.path.join(output_path, "vibe_output_%s.pkl" % person)
os.makedirs(os.path.dirname(dump_path), exist_ok=True)
pickle.dump(vibe_results[person], open(dump_path, 'wb'))
# if not args.no_render:
# # ========= Render results as a single video ========= #
# renderer = Renderer(resolution=(orig_width, orig_height),
# orig_img=True, wireframe=args.wireframe)
# output_img_folder = f'{image_folder}_output'
# os.makedirs(output_img_folder, exist_ok=True)
# print(f'Rendering output video, writing frames to {output_img_folder}')
# # prepare results for rendering
# frame_results = prepare_rendering_results(vibe_results, num_frames)
# mesh_color = {k: colorsys.hsv_to_rgb(
# np.random.rand(), 0.5, 1.0) for k in vibe_results.keys()}
# image_file_names = sorted([
# os.path.join(image_folder, x)
# for x in os.listdir(image_folder)
# if x.endswith('.png') or x.endswith('.jpg')
# ])
# for frame_idx in tqdm(range(len(image_file_names))):
# img_fname = image_file_names[frame_idx]
# img = cv2.imread(img_fname)
# for person_id, person_data in frame_results[frame_idx].items():
# frame_verts = person_data['verts']
# frame_cam = person_data['cam']
# mc = mesh_color[person_id]
# mesh_filename = None
# img = renderer.render(
# img,
# frame_verts,
# cam=frame_cam,
# color=mc,
# mesh_filename=mesh_filename,
# )
# cv2.imwrite(os.path.join(output_img_folder,
# f'{frame_idx:06d}.png'), img)
# if args.display:
# cv2.imshow('Video', img)
# if cv2.waitKey(1) & 0xFF == ord('q'):
# break
# if args.display:
# cv2.destroyAllWindows()
# # ========= Save rendered video ========= #
# vid_name = os.path.basename(video_file)
# save_name = f'{vid_name.replace(".mp4", "")}_vibe_result.mp4'
# save_name = os.path.join(output_path, save_name)
# print(f'Saving result video to {save_name}')
# images_to_video(img_folder=output_img_folder,
# output_vid_file=save_name)
# shutil.rmtree(output_img_folder)
shutil.rmtree(image_folder)
print('================= END =================')
def main(args):
torch.cuda.set_device(args.gpu_id)
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
print(f'Loading video list {args.video_list}')
video_list = [l.strip() for l in open(args.video_list, 'r').readlines()]
if len(video_list) < 1:
print('No files were found in video list')
return
print('Loading VIBE model')
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load VIBE pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
num_videos = len(video_list)
print(f'Processing {num_videos} videos.')
for video_idx, video_file in enumerate(video_list, start=1):
if not osp.isfile(video_file):
print(f'Input video \"{video_file}\" does not exist! Moving on to next file.')
continue
filename = osp.splitext(osp.basename(video_file))[0]
output_path = osp.join(args.output_folder, filename)
os.makedirs(output_path, exist_ok=True)
image_folder, num_frames, img_shape = video_to_images(video_file, return_info=True)
print(f'[{video_idx}/{num_videos}] Processing {num_frames} frames')
orig_height, orig_width = img_shape[:2]
# ========= Run tracking ========= #
bbox_scale = 1.1
if args.tracking_method == 'pose':
if not osp.isabs(video_file):
video_file = osp.join(os.getcwd(), video_file)
tracking_results = run_posetracker(video_file, staf_folder=args.staf_dir, display=args.display)
else:
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = joints2d = None
if args.tracking_method == 'bbox':
bboxes = tracking_results[person_id]['bbox']
elif args.tracking_method == 'pose':
joints2d = tracking_results[person_id]['joints2d']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
dataloader = DataLoader(dataset, batch_size=args.vibe_batch_size, num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:,:,3:75].reshape(batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :,75:].reshape(batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= [Optional] run Temporal SMPLify to refine the results ========= #
if args.run_smplify and args.tracking_method == 'pose':
norm_joints2d = np.concatenate(norm_joints2d, axis=0)
norm_joints2d = convert_kps(norm_joints2d, src='staf', dst='spin')
norm_joints2d = torch.from_numpy(norm_joints2d).float().to(device)
# Run Temporal SMPLify
update, new_opt_vertices, new_opt_cam, new_opt_pose, new_opt_betas, \
new_opt_joints3d, new_opt_joint_loss, opt_joint_loss = smplify_runner(
pred_rotmat=pred_pose,
pred_betas=pred_betas,
pred_cam=pred_cam,
j2d=norm_joints2d,
device=device,
batch_size=norm_joints2d.shape[0],
pose2aa=False,
)
# update the parameters after refinement
print(f'Update ratio after Temporal SMPLify: {update.sum()} / {norm_joints2d.shape[0]}')
pred_verts = pred_verts.cpu()
pred_cam = pred_cam.cpu()
pred_pose = pred_pose.cpu()
pred_betas = pred_betas.cpu()
pred_joints3d = pred_joints3d.cpu()
pred_verts[update] = new_opt_vertices[update]
pred_cam[update] = new_opt_cam[update]
pred_pose[update] = new_opt_pose[update]
pred_betas[update] = new_opt_betas[update]
pred_joints3d[update] = new_opt_joints3d[update]
elif args.run_smplify and args.tracking_method == 'bbox':
print('[WARNING] You need to enable pose tracking to run Temporal SMPLify algorithm!')
print('[WARNING] Continuing without running Temporal SMPLify!..')
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
orig_cam = convert_crop_cam_to_orig_img(
cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height
)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
# Clean-up the temporal folder
# Save the outputs to joblib pkl file. File is loaded through joblib.load(pkl_path)
output_pkl_path = osp.join(args.output_folder, f'{filename}.pkl')
print(f'Saving output results to \"{output_pkl_path}\".')
joblib.dump(vibe_results, output_pkl_path)
if not args.no_render:
# ========= Render results as a single video ========= #
renderer = Renderer(resolution=(orig_width, orig_height), orig_img=True, wireframe=args.wireframe)
output_img_folder = f'{image_folder}_output'
os.makedirs(output_img_folder, exist_ok=True)
print(f'Rendering output video, writing frames to {output_img_folder}')
# prepare results for rendering
frame_results = prepare_rendering_results(vibe_results, num_frames)
mesh_color = {k: colorsys.hsv_to_rgb(np.random.rand(), 0.5, 1.0) for k in vibe_results.keys()}
image_file_names = sorted([
os.path.join(image_folder, x)
for x in os.listdir(image_folder)
if x.endswith('.png') or x.endswith('.jpg')
])
for frame_idx in tqdm(range(len(image_file_names))):
img_fname = image_file_names[frame_idx]
img = cv2.imread(img_fname)
if args.sideview:
side_img = np.zeros_like(img)
for person_id, person_data in frame_results[frame_idx].items():
frame_verts = person_data['verts']
frame_cam = person_data['cam']
mc = mesh_color[person_id]
mesh_filename = None
if args.save_obj:
mesh_folder = os.path.join(output_path, 'meshes', f'{person_id:04d}')
os.makedirs(mesh_folder, exist_ok=True)
mesh_filename = os.path.join(mesh_folder, f'{frame_idx:06d}.obj')
img = renderer.render(
img,
frame_verts,
cam=frame_cam,
color=mc,
mesh_filename=mesh_filename,
)
if args.sideview:
side_img = renderer.render(
side_img,
frame_verts,
cam=frame_cam,
color=mc,
angle=270,
axis=[0,1,0],
)
if args.sideview:
img = np.concatenate([img, side_img], axis=1)
cv2.imwrite(os.path.join(output_img_folder, f'{frame_idx:06d}.png'), img)
if args.display:
cv2.imshow('Video', img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if args.display:
cv2.destroyAllWindows()
# ========= Save rendered video ========= #
vid_name = os.path.basename(video_file)
save_name = f'{vid_name.replace(".mp4", "")}_vibe_result.mp4'
save_name = os.path.join(output_path, save_name)
print(f'Saving result video to {save_name}')
images_to_video(img_folder=output_img_folder, output_vid_file=save_name)
shutil.rmtree(output_img_folder)
# Clean-up after processing
del model
shutil.rmtree(image_folder)
print('================= END =================')
def main(args):
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
video_file = 'test.mp4'
output_folder = 'output/'
output_path = os.path.join(
output_folder,
os.path.basename(video_file).replace('.mp4', ''))
os.makedirs(output_path, exist_ok=True)
print(f'the output path is:{output_path}')
image_folder, num_frames, img_shape = generative()
print(
f'image_folder is:{image_folder},num_frames is:{num_frames},imshape is:{img_shape}'
)
orig_height, orig_width = img_shape[:2]
total_time = time.time()
# ========= Run tracking ========= #
# tracking_method = bbox
bbox_scale = 1.1
# run multi object tracker
mot = MPT(
device=device,
batch_size=12, # 12
display=False, # true
detector_type='yolo', # yolo
output_format='dict',
yolo_img_size=416, #416 * 416
)
tracking_results = mot(image_folder)
print(f'output the result of tracking->{tracking_results}')
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
# 加载预训练模型
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = joints2d = None
# bboxes = n X 4
bboxes = tracking_results[person_id]['bbox']
## frames n X 1
frames = tracking_results[person_id]['frames']
# 一个已经标准化了的图像 dataset 其总体大小为 T X 224 X 224 X 3
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d, # none
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
dataloader = DataLoader(dataset, batch_size=64, num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(
0) #在dim = 0 增加一维张量 变为： 1 X T x 244 x 244 x 3
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2] # 得到 seqlen是一段时间序列图像
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1)) # 1*T X 3
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1)) # 1*T X 72
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1)) # 1*T X 10
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3)) #
pred_cam = torch.cat(pred_cam, dim=0) #(T, 3)
pred_verts = torch.cat(pred_verts, dim=0) #(T, ,3)
pred_pose = torch.cat(pred_pose, dim=0) #(T, 72)
pred_betas = torch.cat(pred_betas, dim=0) #(T, 10)
pred_joints3d = torch.cat(pred_joints3d, dim=0) #(T, ,3)
del batch
# ========= Save results to a pickle file ========= #
# 由tensor 转换成 numpy
pred_cam = pred_cam.cpu().numpy() #(T, 3)
pred_verts = pred_verts.cpu().numpy() #(T, ,3)
pred_pose = pred_pose.cpu().numpy() #(T, 72)
pred_betas = pred_betas.cpu().numpy() #(T, 10)
pred_joints3d = pred_joints3d.cpu().numpy() #(T, ,3)
# 结果为： T X 4
orig_cam = convert_crop_cam_to_orig_img(
cam=pred_cam, # 预测的 cam T X 3
bbox=bboxes, # 目标检测的结果 T X 4
img_width=orig_width, # 最初图像的宽
img_height=orig_height # 最初图像的高
)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
del model
end = time.time()
fps = num_frames / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
total_time = time.time() - total_time
print(
f'Total time spent: {total_time:.2f} seconds (including model loading time).'
)
print(
f'Total FPS (including model loading time): {num_frames / total_time:.2f}.'
)
print(
f'Saving output results to \"{os.path.join(output_path, "vibe_output.pkl")}\".'
)
joblib.dump(vibe_results, os.path.join(output_path, "vibe_output.pkl"))
if not args.no_render:
# ========= Render results as a single video ========= #
# 图像高和宽 resolution
## 一个 工具 真确传参数即可使用
renderer = Renderer(resolution=(orig_width, orig_height),
orig_img=True,
wireframe=args.wireframe)
output_img_folder = f'{image_folder}_output'
os.makedirs(output_img_folder, exist_ok=True)
print(f'Rendering output video, writing frames to {output_img_folder}')
frame_results = prepare_rendering_results(vibe_results, num_frames)
mesh_color = {
k: colorsys.hsv_to_rgb(np.random.rand(), 0.5, 1.0)
for k in vibe_results.keys()
}
image_file_names = sorted([
os.path.join(image_folder, x) for x in os.listdir(image_folder)
if x.endswith('.png') or x.endswith('.jpg')
])
for frame_idx in tqdm(range(len(image_file_names))):
img_fname = image_file_names[frame_idx]
# 读取图片
img = cv2.imread(img_fname)
# true
if args.sideview:
side_img = np.zeros_like(img)
#
for person_id, person_data in frame_results[frame_idx].items():
frame_verts = person_data['verts']
frame_cam = person_data['cam']
mc = mesh_color[person_id]
# 3D模型
mesh_folder = os.path.join(output_path, 'meshes',
f'{person_id:04d}')
os.makedirs(mesh_folder, exist_ok=True)
mesh_filename = os.path.join(mesh_folder,
f'{frame_idx:06d}.obj')
# 输出是一张图片
img = renderer.render(
img,
frame_verts,
cam=frame_cam,
color=mc,
mesh_filename=mesh_filename,
)
side_img = renderer.render(
side_img,
frame_verts,
cam=frame_cam,
color=mc,
angle=270,
axis=[0, 1, 0],
)
img = np.concatenate([img, side_img], axis=1)
cv2.imwrite(
os.path.join(output_img_folder, f'{frame_idx:06d}.png'), img)
if args.display:
cv2.imshow('Video', img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if args.display:
cv2.destroyAllWindows()
print('================= END =================')
def main(args):
if args.device == 'cpu':
device = torch.device('cpu')
print('Running on CPU')
else:
device = torch.device('cuda')
print('Running on GPU')
if args.vid_file:
video_file = args.vid_file
if not os.path.isfile(video_file):
exit(f'Input video \"{video_file}\" does not exist!')
else:
image_file = args.img_file
if not os.path.isfile(image_file):
exit(f'Input video \"{image_file}\" does not exist!')
output_path = os.path.join(
args.output_folder,
os.path.basename(video_file).replace('.mp4', ''))
# output_path = os.path.join(args.output_folder, os.path.basename(video_file).split('.')[0])
os.makedirs(output_path, exist_ok=True)
image_folder, num_frames, img_shape = video_to_images(video_file,
return_info=True)
print(f'Input video number of frames {num_frames}')
orig_height, orig_width = img_shape[:2]
total_time = time.time()
# resize video if too big
# ffmpeg -i input.avi -filter:v scale=720:-1 -c:a copy output.mkv
# ========= Run tracking ========= #
bbox_scale = 1.1
if args.tracking_method == 'pose':
if not os.path.isabs(video_file):
video_file = os.path.join(os.getcwd(), video_file)
tracking_results = run_posetracker(video_file,
staf_folder=args.staf_dir,
display=args.display)
else:
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
import pdb
pdb.set_trace
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=True)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = joints2d = None
if args.tracking_method == 'bbox':
bboxes = tracking_results[person_id]['bbox']
elif args.tracking_method == 'pose':
joints2d = tracking_results[person_id]['joints2d']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
dataloader = DataLoader(dataset,
batch_size=args.vibe_batch_size,
num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= [Optional] run Temporal SMPLify to refine the results ========= #
if args.run_smplify and args.tracking_method == 'pose':
norm_joints2d = np.concatenate(norm_joints2d, axis=0)
norm_joints2d = convert_kps(norm_joints2d, src='staf', dst='spin')
norm_joints2d = torch.from_numpy(norm_joints2d).float().to(device)
# Run Temporal SMPLify
update, new_opt_vertices, new_opt_cam, new_opt_pose, new_opt_betas, \
new_opt_joints3d, new_opt_joint_loss, opt_joint_loss = smplify_runner(
pred_rotmat=pred_pose,
pred_betas=pred_betas,
pred_cam=pred_cam,
j2d=norm_joints2d,
device=device,
batch_size=norm_joints2d.shape[0],
pose2aa=False,
)
# update the parameters after refinement
print(
f'Update ratio after Temporal SMPLify: {update.sum()} / {norm_joints2d.shape[0]}'
)
pred_verts = pred_verts.cpu()
pred_cam = pred_cam.cpu()
pred_pose = pred_pose.cpu()
pred_betas = pred_betas.cpu()
pred_joints3d = pred_joints3d.cpu()
pred_verts[update] = new_opt_vertices[update]
pred_cam[update] = new_opt_cam[update]
pred_pose[update] = new_opt_pose[update]
pred_betas[update] = new_opt_betas[update]
pred_joints3d[update] = new_opt_joints3d[update]
elif args.run_smplify and args.tracking_method == 'bbox':
print(
'[WARNING] You need to enable pose tracking to run Temporal SMPLify algorithm!'
)
print('[WARNING] Continuing without running Temporal SMPLify!..')
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
del model
end = time.time()
fps = num_frames / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
total_time = time.time() - total_time
print(
f'Total time spent: {total_time:.2f} seconds (including model loading time).'
)
print(
f'Total FPS (including model loading time): {num_frames / total_time:.2f}.'
)
print(
f'Saving output results to \"{os.path.join(output_path, "vibe_output.pkl")}\".'
)
joblib.dump(vibe_results, os.path.join(output_path, "vibe_output.pkl"))
if not args.no_render:
# ========= Render results as a single video ========= #
renderer = Renderer(resolution=(orig_width, orig_height),
orig_img=True,
wireframe=args.wireframe)
output_img_folder = f'{image_folder}_images'
os.makedirs(output_img_folder, exist_ok=True)
print(f'Rendering output video, writing frames to {output_img_folder}')
output_pose_folder = f'{image_folder}_poses'
os.makedirs(output_pose_folder, exist_ok=True)
print(f'Saving poses to {output_pose_folder}')
# prepare results for rendering
from numpy import save
save(f'{os.path.basename(video_file)}_poses.npy',
vibe_results[1]['joints3d'][:, :25, :])
print('Saving numpy poses file to' + f'{video_file}_poses.npy')
frame_results = prepare_rendering_results(
vibe_results,
num_frames) # returns a list of dicts (one dict for each person)
mesh_color = {
k: colorsys.hsv_to_rgb(np.random.rand(), 0.5, 1.0)
for k in vibe_results.keys()
}
image_file_names = sorted([
os.path.join(image_folder, x) for x in os.listdir(image_folder)
if x.endswith('.png') or x.endswith('.jpg')
])
for frame_idx in tqdm(range(len(image_file_names))):
img_fname = image_file_names[frame_idx]
img = cv2.imread(img_fname)
if args.sideview:
side_img = np.zeros_like(img)
for person_id, person_data in frame_results[frame_idx].items():
frame_verts = person_data['verts']
frame_cam = person_data['cam']
frame_pose = person_data['joints3d'][:25]
mc = mesh_color[person_id]
mesh_filename = None
if args.save_obj:
mesh_folder = os.path.join(output_path, 'meshes',
f'{person_id:04d}')
os.makedirs(mesh_folder, exist_ok=True)
mesh_filename = os.path.join(mesh_folder,
f'{frame_idx:06d}.obj')
# bgr image (opencv format)
img = renderer.render(
img,
frame_verts,
cam=frame_cam,
color=mc,
mesh_filename=mesh_filename,
)
# import pdb; pdb.set_trace()
# Create a 3D projection and save as img
# pose is mirrored
# plot_skeleton(output_pose_folder, frame_idx, frame_pose)
if args.sideview:
side_img = renderer.render(
side_img,
frame_verts,
cam=frame_cam,
color=mc,
angle=270,
axis=[0, 1, 0],
)
if args.sideview:
img = np.concatenate([img, side_img], axis=1)
# concatenate pose img with this image before writing
cv2.imwrite(
os.path.join(output_img_folder, f'{frame_idx:06d}.png'), img)
if args.display:
cv2.imshow('Video', img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if args.display:
cv2.destroyAllWindows()
# ========= Save rendered video ========= #
vid_name = os.path.basename(video_file)
save_name = f'{vid_name.replace(".mp4", "")}_vibe_result.mp4'
save_name = os.path.join(output_path, save_name)
print(f'Saving result video to {save_name}')
images_to_video(img_folder=output_img_folder,
output_vid_file=save_name)
# shutil.rmtree(output_img_folder)
shutil.rmtree(image_folder)
print('================= END =================')
def inferData(self, image_path, output_path, pretrained):
img0 = cv2.imread(osp.join(image_path, os.listdir(image_path)[0]))
num_frames = len(os.listdir(image_path))
img_height, img_width = img0.shape[0:2]
tracking_start_time = time.time()
tracking_results = self.Tracker(image_path)
tracking_end_time = time.time()
print(
f'Tracking fps: {(num_frames)/(tracking_end_time - tracking_start_time): .2f}'
)
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < 25:
del tracking_results[person_id]
ckpt = torch.load(pretrained)
ckpt = ckpt['gen_state_dict']
self.generator.load_state_dict(ckpt, strict=False)
self.generator.eval()
vibe_results = {}
vibeFPS = 0
for person_id in tqdm(list(tracking_results.keys())):
person_start_time = time.time()
joints2d = None
bboxes = tracking_results[person_id]['bbox']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_path,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=1.0,
)
bboxes = dataset.bboxes
frames = dataset.frames
dataloader = DataLoader(dataset, batch_size=12, num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
batch = batch.unsqueeze(0)
batch = batch.cuda()
batch_size, seqlen = batch.shape[:2]
# output = model(batch)[-1]
output = self.generator(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
person_end_time = time.time()
person_time = person_end_time - person_start_time
person_frame = len(frames)
vibeFPS += person_frame / person_time
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=img_width,
img_height=img_height)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
joblib.dump(vibe_results, os.path.join(output_path, "vibe_output.pkl"))
print(f'Mean target fps: {vibeFPS/len(list(tracking_results.keys()))}')
return vibe_results
def main(args):
# check GPU availability
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
# get input video
video_file = args.input_video
# save hand info
save_hand_csv = args.save_hand_csv
# ========= [Optional] download the youtube video ========= #
if video_file.startswith('https://www.youtube.com'):
print(f'Donwloading YouTube video \"{video_file}\"')
video_file = download_youtube_clip(video_file, '/tmp')
if video_file is None:
exit('Youtube url is not valid!')
print(f'YouTube Video has been downloaded to {video_file}...')
# check video existence
if not os.path.isfile(video_file):
exit(f'Input video \"{video_file}\" does not exist!')
# set output flles
output_path = args.output_folder
image_folder, num_frames, img_shape = video_to_images(video_file, \
"/tmp/" + output_path.split("/")[-1], return_info=True)
print(f'Input video number of frames {num_frames}')
orig_height, orig_width = img_shape[:2]
# get the frame rate (frames per second) of the input video
video = cv2.VideoCapture(video_file)
fps = video.get(cv2.CAP_PROP_FPS)
# ========= Run tracking ========= #
bbox_scale = 1.1
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
# only focus on the frame with the first person
largest_num_frames = 0
largest_person = None
for person_id in list(tracking_results.keys()):
num_frames = tracking_results[person_id]['frames'].shape[0]
if num_frames <= largest_num_frames:
del tracking_results[person_id]
else:
largest_num_frames = tracking_results[person_id]['frames'].shape[0]
if largest_person != None:
del tracking_results[largest_person]
largest_person = person_id
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = tracking_results[person_id]['bbox']
frames = tracking_results[person_id]['frames']
np.save(output_path + "/frames", frames)
# inference data of each person
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
# load data
dataloader = DataLoader(dataset,
batch_size=args.vibe_batch_size,
num_workers=16)
# extract data
with torch.no_grad():
pred_cam, pred_verts, pred_pose = [], [], []
for batch in dataloader:
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
del batch
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
# get camera pose
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
# get result information
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'bboxes': bboxes,
'frame_ids': frames,
}
np.savetxt(output_path + "/pred_cam.csv", pred_cam, delimiter=",")
np.savetxt(output_path + "/orig_cam.csv", orig_cam, delimiter=",")
vibe_results[person_id] = output_dict
del model
frame_results = prepare_rendering_results(vibe_results, len(frames))
np.save(output_path + "/frame_results", frame_results)
np.save(output_path + "/image_folder", image_folder)
np.save(output_path + "/orig_width", orig_width)
np.save(output_path + "/orig_height", orig_height)
def runDemo(image_folder, output_folder, pretrained, tracker_batch_size=12, vibe_batch_size=450, wireframe=False):
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
output_path = os.path.join(output_folder, os.path.basename(image_folder).replace('.mp4', ''))
os.makedirs(output_path, exist_ok=True)
num_frames, img_shape = img_folder_Info(image_folder)
print(f'Input video number of frames {num_frames}')
orig_height, orig_width = img_shape[:2]
total_time = time.time()
# ========= Run tracking ========= #
bbox_scale = 1.0
# run multi object tracker
mot = MPT(
device=device,
batch_size=tracker_batch_size,
display=False,
detector_type='yolo',
output_format='dict',
yolo_img_size=416,
)
tracking_results = mot(image_folder)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Define VIBE model ========= #
model = e2e_VIBE(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
pretrained_file = pretrained
ckpt = torch.load(pretrained_file)
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
time_results = {}
for person_id in tqdm(list(tracking_results.keys())):
person_start_time = time.time()
joints2d = None
bboxes = tracking_results[person_id]['bbox']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
dataloader = DataLoader(dataset, batch_size=vibe_batch_size, num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
# output = model(batch, J_regressor=J_regressor)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:,:,3:75].reshape(batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :,75:].reshape(batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
person_end_time = time.time()
person_time = person_end_time - person_start_time
person_frame = len(frames)
print(f'Person Time: {person_time:.2f}, Person FPS:{person_frame/person_time: .2f} ')
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
orig_cam = convert_crop_cam_to_orig_img(
cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height
)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
del model
end = time.time()
fps = num_frames / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
total_time = time.time() - total_time
print(f'Total time spent: {total_time:.2f} seconds (including model loading time).')
print(f'Total FPS (including model loading time): {num_frames / total_time:.2f}.')
print(f'Saving output results to \"{os.path.join(output_path, "vibe_output.pkl")}\".')
joblib.dump(vibe_results, os.path.join(output_folder, "vibe_output.pkl"))
# ========= Render results as a single video ========= #
renderer = Renderer(resolution=(orig_width, orig_height), orig_img=True, wireframe=wireframe)
output_img_folder = f'{image_folder}_output'
os.makedirs(output_img_folder, exist_ok=True)
print(f'Rendering output video, writing frames to {output_img_folder}')
# prepare results for rendering
frame_results = prepare_rendering_results(vibe_results, num_frames)
mesh_color = {k: colorsys.hsv_to_rgb(np.random.rand(), 0.5, 1.0) for k in vibe_results.keys()}
image_file_names = sorted([
os.path.join(image_folder, x)
for x in os.listdir(image_folder)
if x.endswith('.png') or x.endswith('.jpg')
])
for frame_idx in tqdm(range(len(image_file_names))):
img_fname = image_file_names[frame_idx]
img = cv2.imread(img_fname)
for person_id, person_data in frame_results[frame_idx].items():
frame_verts = person_data['verts']
frame_cam = person_data['cam']
mc = mesh_color[person_id]
mesh_filename = None
img = renderer.render(
img,
frame_verts,
cam=frame_cam,
color=mc,
mesh_filename=mesh_filename,
)
cv2.imwrite(os.path.join(output_img_folder, f'{frame_idx:06d}.png'), img)
# ========= Save rendered video ========= #
vid_name = os.path.basename(image_folder)
save_name = 'vibe_result.mp4'
save_name = os.path.join(output_folder, save_name)
print(f'Saving result video to {save_name}')
images_to_video(img_folder=output_img_folder, output_vid_file=save_name)
shutil.rmtree(output_img_folder)
print('================= END =================')
def detect_track_vibe(frame_orig, model_dettr, model_vibe, device):
orig_height, orig_width = frame_orig.shape[:2]
frame = cv2.cvtColor(frame_orig, cv2.COLOR_BGR2RGB)
frame = frame / 255.
frame = frame.transpose((2, 0, 1))
frame = torch.from_numpy(frame)
frame = frame.unsqueeze(0)
tracking_results = model_dettr(frame)
# print('1111111111111tracking result',tracking_results)
# print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in list(tracking_results.keys()):
bboxes = joints2d = None
bboxes = tracking_results[person_id]['bbox'] # shape(1,4)
# print('bboxes: ',bboxes) #相同
frames = tracking_results[person_id]['frames']
# print('22222222',bboxes)
bbox_scale = 1.1
dataset = Inference(frame=frame_orig, bboxes=bboxes, scale=bbox_scale)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
batch = dataset
batch = batch.unsqueeze(0).unsqueeze(0)
batch = batch.to(device)
# print(batch.shape)
batch_size, seqlen = batch.shape[:2]
output = model_vibe(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(batch_size * seqlen, -1,
3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
pred_cam = pred_cam.cpu().numpy()
# print('pred_cam: ',pred_cam) #不同
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
# print('3333333333',pred_cam)
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
# print('orig_cam',orig_cam.shape)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
# print('vibe_results orig_cam: ',vibe_results[1]['orig_cam'])
# print('vibe_results pose: ', vibe_results[1]['pose'])
end = time.time()
fps = 1 / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
return vibe_results
def main(args):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Configure depth and color streams
pipeline = rs.pipeline()
config = rs.config()
config.enable_stream(rs.stream.depth, 640, 480, rs.format.z16, 30)
config.enable_stream(rs.stream.color, 640, 480, rs.format.bgr8, 30)
# Start streaming
pipeline.start(config)
# ========= Run tracking ========= #
bbox_scale = 1.1
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
# ========= main loop ======================= #
fourcc = cv2.VideoWriter_fourcc(*'XVID')
#cap = cv2.VideoCapture('test.avi')
#out = cv2.VideoWriter('output/test_5people.avi', fourcc, 30.0, (640, 360), True)
#cap = cv2.VideoCapture('sample_video.mp4')
out = cv2.VideoWriter('output/test_realsense_2.avi', fourcc, 10.0,
(640, 480), True)
# load renderer
renderer = Renderer(resolution=(640, 480),
orig_img=True,
wireframe=args.wireframe)
i = 0
time_acc = 0.0
while (True):
# Capture frame-by-frame
total_time = time.time()
frames = pipeline.wait_for_frames()
frame_orig = frames.get_color_frame()
# Convert images to numpy arrays
frame_orig = np.asanyarray(frame_orig.get_data())
#ret, frame_orig = cap.read()
if frame_orig is None:
break
#for i in range(1,300):
# total_time = time.time()
# path = os.path.join('tmp/sample_video/',f'{i:06d}.png')
# frame_orig = cv2.imread(path)
orig_height, orig_width = frame_orig.shape[:2]
frame = cv2.cvtColor(frame_orig, cv2.COLOR_BGR2RGB)
frame = frame / 255.
frame = frame.transpose((2, 0, 1))
frame = torch.from_numpy(frame)
frame = frame.unsqueeze(0)
tracking_results = mot(frame)
#print('1111111111111tracking result',tracking_results)
#print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in list(tracking_results.keys()):
bboxes = joints2d = None
bboxes = tracking_results[person_id]['bbox'] # shape(1,4)
#print('bboxes: ',bboxes) #相同
frames = tracking_results[person_id]['frames']
#print('22222222',bboxes)
dataset = Inference(frame=frame_orig,
bboxes=bboxes,
scale=bbox_scale)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
batch = dataset
batch = batch.unsqueeze(0).unsqueeze(0)
batch = batch.to(device)
#print(batch.shape)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
pred_cam = pred_cam.cpu().numpy()
#print('pred_cam: ',pred_cam) #不同
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
#print('3333333333',pred_cam)
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
#print('orig_cam',orig_cam.shape)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
#print('vibe_results orig_cam: ',vibe_results[1]['orig_cam'])
#print('vibe_results pose: ', vibe_results[1]['pose'])
end = time.time()
fps = 1 / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
if not args.no_render:
render_time = time.time()
# load renderer
#renderer = Renderer(resolution=(orig_width, orig_height), orig_img=True, wireframe=args.wireframe)
# prepare results for rendering
num_frames = 1
#print('vibe_results1111: ',vibe_results)
#vibe_results[1]['orig_cam'] = vibe_results[1]['orig_cam'][np.newaxis,:]
#print('orig_cam: ',vibe_results[1]['orig_cam'].shape)
frame_results = prepare_rendering_results(vibe_results, num_frames)
#print('frame_results',frame_results)
img = frame_orig
mesh_color = {
k: colorsys.hsv_to_rgb(np.random.rand(), 0.5, 1.0)
for k in vibe_results.keys()
}
#img = frame
if args.sideview:
side_img = np.zeros_like(img)
for person_id, person_data in frame_results[0].items():
frame_verts = person_data['verts']
frame_cam = person_data['cam']
#print('4444444444frame_cam',frame_cam)
mc = mesh_color[person_id]
mesh_filename = None
img = renderer.render(
img,
frame_verts,
cam=frame_cam,
color=mc,
mesh_filename=mesh_filename,
)
fps = 1 / (time.time() - render_time)
print(f'RENDER FPS: {fps:.2f}')
#img = img.numpy()
out.write(img)
if args.display:
cv2.imshow('Video', img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if args.display:
cv2.destroyAllWindows()
total_time = time.time() - total_time
i += 1
time_acc += total_time
print(
'num of frame: ', i,
f' Total time spent: {total_time:.2f} seconds (detect+track+vibe+render).'
)
print(f'FPS : { 1 / total_time:.2f}.')
print('Total average FPS: ', i / time_acc)
# ========= Save rendered video ========= #
print('================= END =================')
def main(args):
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
dict = {
'bridge': 1,
'childs': 2,
'downwarddog': 3,
'mountain': 4,
'plank': 5,
'seatedforwardbend': 6,
'tree': 7,
'trianglepose': 8,
'warrior1': 9,
'warrior2': 10
}
dir_path = '/home/ubuntu/PoseEstimation/VIBE/InputData/input_test_set/'
output_folder = '/home/ubuntu/PoseEstimation/VIBE/OutputData/test_set/'
joints3D_csv = open('output_joints3d_dog.csv', 'a')
pose_csv = open('output_pose.csv_dog', 'a')
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
video_file = '/home/ubuntu/PoseEstimation/VIBE/DogVideo.mp4'
video_label = dict['bridge']
if not os.path.isfile(video_file):
exit(f'Input video \"{video_file}\" does not exist!')
image_folder, num_frames, img_shape = video_to_images(video_file,
return_info=True)
print(f'Input video number of frames {num_frames}')
orig_height, orig_width = img_shape[:2]
total_time = time.time()
# ========= Run tracking ========= #
bbox_scale = 1.1
if args.tracking_method == 'pose':
if not os.path.isabs(video_file):
video_file = os.path.join(os.getcwd(), video_file)
tracking_results = run_posetracker(video_file,
staf_folder=args.staf_dir,
display=args.display)
else:
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = joints2d = None
if args.tracking_method == 'bbox':
bboxes = tracking_results[person_id]['bbox']
elif args.tracking_method == 'pose':
joints2d = tracking_results[person_id]['joints2d']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
dataloader = DataLoader(dataset,
batch_size=args.vibe_batch_size,
num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= [Optional] run Temporal SMPLify to refine the results ========= #
if args.run_smplify and args.tracking_method == 'pose':
norm_joints2d = np.concatenate(norm_joints2d, axis=0)
norm_joints2d = convert_kps(norm_joints2d, src='staf', dst='spin')
norm_joints2d = torch.from_numpy(norm_joints2d).float().to(device)
# Run Temporal SMPLify
update, new_opt_vertices, new_opt_cam, new_opt_pose, new_opt_betas, \
new_opt_joints3d, new_opt_joint_loss, opt_joint_loss = smplify_runner(
pred_rotmat=pred_pose,
pred_betas=pred_betas,
pred_cam=pred_cam,
j2d=norm_joints2d,
device=device,
batch_size=norm_joints2d.shape[0],
pose2aa=False,
)
# update the parameters after refinement
print(
f'Update ratio after Temporal SMPLify: {update.sum()} / {norm_joints2d.shape[0]}'
)
pred_verts = pred_verts.cpu()
pred_cam = pred_cam.cpu()
pred_pose = pred_pose.cpu()
pred_betas = pred_betas.cpu()
pred_joints3d = pred_joints3d.cpu()
pred_verts[update] = new_opt_vertices[update]
pred_cam[update] = new_opt_cam[update]
pred_pose[update] = new_opt_pose[update]
pred_betas[update] = new_opt_betas[update]
pred_joints3d[update] = new_opt_joints3d[update]
elif args.run_smplify and args.tracking_method == 'bbox':
print(
'[WARNING] You need to enable pose tracking to run Temporal SMPLify algorithm!'
)
print('[WARNING] Continuing without running Temporal SMPLify!..')
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
# Runs 1 Euro Filter to smooth out the results
if args.smooth:
min_cutoff = args.smooth_min_cutoff # 0.004
beta = args.smooth_beta # 1.5
print(
f'Running smoothing on person {person_id}, min_cutoff: {min_cutoff}, beta: {beta}'
)
pred_verts, pred_pose, pred_joints3d = smooth_pose(
pred_pose, pred_betas, min_cutoff=min_cutoff, beta=beta)
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
for i in range(len(output_dict['joints3d'])):
if (i % 5 == 0):
flat_arr = output_dict['joints3d'][i].flatten()
len_N = len(flat_arr)
np.savetxt(joints3D_csv, [np.append(flat_arr, [video_label])],
delimiter=',',
fmt=' '.join(['%f'] * len_N + ['%i']))
for i in range(len(output_dict['pose'])):
if (i % 5 == 0):
pose_arr = output_dict['pose'][i].flatten()
len_M = len(pose_arr)
np.savetxt(pose_csv, [np.append(pose_arr, [video_label])],
delimiter=',',
fmt=' '.join(['%f'] * len_M + ['%i']))
end = time.time()
fps = num_frames / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
total_time = time.time() - total_time
print(
f'Total time spent: {total_time:.2f} seconds (including model loading time).'
)
print(
f'Total FPS (including model loading time): {num_frames / total_time:.2f}.'
)
def main(args):
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
""" Prepare input video (images) """
video_file = args.vid_file
if video_file.startswith('https://www.youtube.com'):
print(f"Donwloading YouTube video \'{video_file}\'")
video_file = download_youtube_clip(video_file, '/tmp')
if video_file is None:
exit('Youtube url is not valid!')
print(f"YouTube Video has been downloaded to {video_file}...")
if not os.path.isfile(video_file):
exit(f"Input video \'{video_file}\' does not exist!")
output_path = osp.join('./output/demo_output',
os.path.basename(video_file).replace('.mp4', ''))
Path(output_path).mkdir(parents=True, exist_ok=True)
image_folder, num_frames, img_shape = video_to_images(video_file,
return_info=True)
print(f"Input video number of frames {num_frames}\n")
orig_height, orig_width = img_shape[:2]
""" Run tracking """
total_time = time.time()
bbox_scale = 1.2
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
""" Get TCMR model """
seq_len = 16
model = TCMR(seqlen=seq_len, n_layers=2, hidden_size=1024).to(device)
# Load pretrained weights
pretrained_file = args.model
ckpt = torch.load(pretrained_file)
print(f"Load pretrained weights from \'{pretrained_file}\'")
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
# Change mesh gender
gender = args.gender # 'neutral', 'male', 'female'
model.regressor.smpl = SMPL(SMPL_MODEL_DIR,
batch_size=64,
create_transl=False,
gender=gender).cuda()
model.eval()
# Get feature_extractor
from lib.models.spin import hmr
hmr = hmr().to(device)
checkpoint = torch.load(
osp.join(BASE_DATA_DIR, 'spin_model_checkpoint.pth.tar'))
hmr.load_state_dict(checkpoint['model'], strict=False)
hmr.eval()
""" Run TCMR on each person """
print("\nRunning TCMR on each person tracklet...")
tcmr_time = time.time()
tcmr_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = joints2d = None
bboxes = tracking_results[person_id]['bbox']
frames = tracking_results[person_id]['frames']
# Prepare static image features
dataset = CropDataset(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
crop_dataloader = DataLoader(dataset, batch_size=256, num_workers=16)
with torch.no_grad():
feature_list = []
for i, batch in enumerate(crop_dataloader):
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.to(device)
feature = hmr.feature_extractor(batch.reshape(-1, 3, 224, 224))
feature_list.append(feature.cpu())
del batch
feature_list = torch.cat(feature_list, dim=0)
# Encode temporal features and estimate 3D human mesh
dataset = FeatureDataset(
image_folder=image_folder,
frames=frames,
seq_len=seq_len,
)
dataset.feature_list = feature_list
dataloader = DataLoader(dataset, batch_size=64, num_workers=32)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for i, batch in enumerate(dataloader):
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.to(device)
output = model(batch)[0][-1]
pred_cam.append(output['theta'][:, :3])
pred_verts.append(output['verts'])
pred_pose.append(output['theta'][:, 3:75])
pred_betas.append(output['theta'][:, 75:])
pred_joints3d.append(output['kp_3d'])
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
bboxes[:, 2:] = bboxes[:, 2:] * 1.2
if args.render_plain:
pred_cam[:, 0], pred_cam[:, 1:] = 1, 0 # np.array([[1, 0, 0]])
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
tcmr_results[person_id] = output_dict
del model
end = time.time()
fps = num_frames / (end - tcmr_time)
print(f'TCMR FPS: {fps:.2f}')
total_time = time.time() - total_time
print(
f'Total time spent: {total_time:.2f} seconds (including model loading time).'
)
print(
f'Total FPS (including model loading time): {num_frames / total_time:.2f}.'
)
if args.save_pkl:
print(
f"Saving output results to \'{os.path.join(output_path, 'tcmr_output.pkl')}\'."
)
joblib.dump(tcmr_results, os.path.join(output_path, "tcmr_output.pkl"))
""" Render results as a single video """
renderer = Renderer(resolution=(orig_width, orig_height),
orig_img=True,
wireframe=args.wireframe)
output_img_folder = f'{image_folder}_output'
input_img_folder = f'{image_folder}_input'
os.makedirs(output_img_folder, exist_ok=True)
os.makedirs(input_img_folder, exist_ok=True)
print(f"\nRendering output video, writing frames to {output_img_folder}")
# prepare results for rendering
frame_results = prepare_rendering_results(tcmr_results, num_frames)
mesh_color = {
k: colorsys.hsv_to_rgb(np.random.rand(), 0.5, 1.0)
for k in tcmr_results.keys()
}
image_file_names = sorted([
os.path.join(image_folder, x) for x in os.listdir(image_folder)
if x.endswith('.png') or x.endswith('.jpg')
])
for frame_idx in tqdm(range(len(image_file_names))):
img_fname = image_file_names[frame_idx]
img = cv2.imread(img_fname)
input_img = img.copy()
if args.render_plain:
img[:] = 0
if args.sideview:
side_img = np.zeros_like(img)
for person_id, person_data in frame_results[frame_idx].items():
frame_verts = person_data['verts']
frame_cam = person_data['cam']
mesh_filename = None
if args.save_obj:
mesh_folder = os.path.join(output_path, 'meshes',
f'{person_id:04d}')
Path(mesh_folder).mkdir(parents=True, exist_ok=True)
mesh_filename = os.path.join(mesh_folder,
f'{frame_idx:06d}.obj')
mc = mesh_color[person_id]
img = renderer.render(
img,
frame_verts,
cam=frame_cam,
color=mc,
mesh_filename=mesh_filename,
)
if args.sideview:
side_img = renderer.render(
side_img,
frame_verts,
cam=frame_cam,
color=mc,
angle=270,
axis=[0, 1, 0],
)
if args.sideview:
img = np.concatenate([img, side_img], axis=1)
# save output frames
cv2.imwrite(os.path.join(output_img_folder, f'{frame_idx:06d}.jpg'),
img)
cv2.imwrite(os.path.join(input_img_folder, f'{frame_idx:06d}.jpg'),
input_img)
if args.display:
cv2.imshow('Video', img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if args.display:
cv2.destroyAllWindows()
""" Save rendered video """
vid_name = os.path.basename(video_file)
save_name = f'tcmr_{vid_name.replace(".mp4", "")}_output.mp4'
save_path = os.path.join(output_path, save_name)
images_to_video(img_folder=output_img_folder, output_vid_file=save_path)
images_to_video(img_folder=input_img_folder,
output_vid_file=os.path.join(output_path, vid_name))
print(f"Saving result video to {os.path.abspath(save_path)}")
shutil.rmtree(output_img_folder)
shutil.rmtree(input_img_folder)
shutil.rmtree(image_folder)
def main(args):
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
video_file = args.vid_file
# ========= [Optional] download the youtube video ========= #
if video_file.startswith('https://www.youtube.com'):
print(f'Donwloading YouTube video \"{video_file}\"')
video_file = download_youtube_clip(video_file, '/tmp')
if video_file is None:
exit('Youtube url is not valid!')
print(f'YouTube Video has been downloaded to {video_file}...')
if not os.path.isfile(video_file):
exit(f'Input video \"{video_file}\" does not exist!')
output_path = os.path.join(args.output_folder, os.path.basename(video_file).replace('.mp4', ''))
os.makedirs(output_path, exist_ok=True)
image_folder, num_frames, img_shape = video_to_images(video_file, return_info=True)
print(f'Input video number of frames {num_frames}')
orig_height, orig_width = img_shape[:2]
total_time = time.time()
# ========= Run tracking ========= #
bbox_scale = 1.1
if args.tracking_method == 'pose':
if not os.path.isabs(video_file):
video_file = os.path.join(os.getcwd(), video_file)
tracking_results = run_posetracker(video_file, staf_folder=args.staf_dir, display=args.display)
else:
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = joints2d = None
if args.tracking_method == 'bbox':
bboxes = tracking_results[person_id]['bbox']
elif args.tracking_method == 'pose':
joints2d = tracking_results[person_id]['joints2d']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
# reduce the num of worker if you encountered the error: DLL load failed: The paging file is too small for this operation to complete
dataloader = DataLoader(dataset, batch_size=args.vibe_batch_size, num_workers=8)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:,:,3:75].reshape(batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :,75:].reshape(batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= [Optional] run Temporal SMPLify to refine the results ========= #
if args.run_smplify and args.tracking_method == 'pose':
norm_joints2d = np.concatenate(norm_joints2d, axis=0)
norm_joints2d = convert_kps(norm_joints2d, src='staf', dst='spin')
norm_joints2d = torch.from_numpy(norm_joints2d).float().to(device)
# Run Temporal SMPLify
update, new_opt_vertices, new_opt_cam, new_opt_pose, new_opt_betas, \
new_opt_joints3d, new_opt_joint_loss, opt_joint_loss = smplify_runner(
pred_rotmat=pred_pose,
pred_betas=pred_betas,
pred_cam=pred_cam,
j2d=norm_joints2d,
device=device,
batch_size=norm_joints2d.shape[0],
pose2aa=False,
)
# update the parameters after refinement
print(f'Update ratio after Temporal SMPLify: {update.sum()} / {norm_joints2d.shape[0]}')
pred_verts = pred_verts.cpu()
pred_cam = pred_cam.cpu()
pred_pose = pred_pose.cpu()
pred_betas = pred_betas.cpu()
pred_joints3d = pred_joints3d.cpu()
pred_verts[update] = new_opt_vertices[update]
pred_cam[update] = new_opt_cam[update]
pred_pose[update] = new_opt_pose[update]
pred_betas[update] = new_opt_betas[update]
pred_joints3d[update] = new_opt_joints3d[update]
elif args.run_smplify and args.tracking_method == 'bbox':
print('[WARNING] You need to enable pose tracking to run Temporal SMPLify algorithm!')
print('[WARNING] Continuing without running Temporal SMPLify!..')
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
orig_cam = convert_crop_cam_to_orig_img(
cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height
)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
del model
end = time.time()
fps = num_frames / (end - vibe_time)
print(f'VIBE FPS: {fps:.2f}')
total_time = time.time() - total_time
print(f'Total time spent: {total_time:.2f} seconds (including model loading time).')
print(f'Total FPS (including model loading time): {num_frames / total_time:.2f}.')
print(f'Saving output results to \"{os.path.join(output_path, "vibe_output.pkl")}\".')
joblib.dump(vibe_results, os.path.join(output_path, "vibe_output.pkl"))
if not args.no_render:
# ========= Render results as a single video ========= #
renderer = Renderer(resolution=(orig_width, orig_height), orig_img=True, wireframe=args.wireframe)
output_img_folder = f'{image_folder}_output'
os.makedirs(output_img_folder, exist_ok=True)
print(f'Rendering output video, writing frames to {output_img_folder}')
# prepare results for rendering
frame_results = prepare_rendering_results(vibe_results, num_frames)
mesh_color = {k: colorsys.hsv_to_rgb(np.random.rand(), 0.5, 1.0) for k in vibe_results.keys()}
image_file_names = sorted([
os.path.join(image_folder, x)
for x in os.listdir(image_folder)
if x.endswith('.png') or x.endswith('.jpg')
])
for frame_idx in tqdm(range(len(image_file_names))):
img_fname = image_file_names[frame_idx]
img = cv2.imread(img_fname)
if args.sideview:
side_img = np.zeros_like(img)
for person_id, person_data in frame_results[frame_idx].items():
frame_verts = person_data['verts']
frame_cam = person_data['cam']
mc = mesh_color[person_id]
mesh_filename = None
if args.save_obj:
mesh_folder = os.path.join(output_path, 'meshes', f'{person_id:04d}')
os.makedirs(mesh_folder, exist_ok=True)
mesh_filename = os.path.join(mesh_folder, f'{frame_idx:06d}.obj')
img = renderer.render(
img,
frame_verts,
cam=frame_cam,
color=mc,
mesh_filename=mesh_filename,
)
if args.sideview:
side_img = renderer.render(
side_img,
frame_verts,
cam=frame_cam,
color=mc,
angle=270,
axis=[0,1,0],
)
if args.sideview:
img = np.concatenate([img, side_img], axis=1)
font = cv2.FONT_HERSHEY_SIMPLEX
x = 10 #position of text
y = 20 #position of text
cv2.putText(img, str(frame_idx), (x,y), font ,0.55,(0,255,0),1)
cv2.imwrite(os.path.join(output_img_folder, f'{frame_idx:06d}.png'), img)
if args.display:
cv2.imshow('Video', img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if args.display:
cv2.destroyAllWindows()
# ========= Save rendered video ========= #
vid_name = os.path.basename(video_file)
save_name = f'{vid_name.replace(".mp4", "")}_vibe_result.mp4'
save_name = os.path.join(output_path, save_name)
print(f'Saving result video to {save_name}')
images_to_video(img_folder=output_img_folder, output_vid_file=save_name)
shutil.rmtree(output_img_folder)
shutil.rmtree(image_folder)
# generate and save the joints csv file for animating avatars later
output = joblib.load(os.path.join(output_path, "vibe_output.pkl"))
for i in output.keys():
print('Track ids:', i , end='\n\n')
num_ppl = len(output.keys())
print('VIBE output file content:', end='\n\n')
vid_name = os.path.basename(video_file)
vibe_result_folder = output_path
# output the pose result as csv
# format: v_personId_numFrames
pose_filename_list = []
for i in output.keys():
pose_filename = vibe_result_folder + "/" + vid_name + "_"+ str(i) + "_" + str(output[i]['pose'].shape[0]) + ".csv"
pose_filename_list.append(pose_filename)
field_names = []
for idx in range(73): # 72 -> 73 (+ frame_id at 0)
field_names.append(str(idx))
with open(pose_filename, 'w', newline='') as file:
writer = csv.writer(file)
writer.writerow(field_names)
for frame_id in range(len(output[i]['pose'])):
output_data = [output[i]['frame_ids'][frame_id]]
output_data.extend(output[i]['pose'][frame_id])
#print(output_data)
writer.writerow(output_data)
print('================= END =================')
def main(args):
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
map_vals = {
'bridge': 1,
'childs': 2,
'downwarddog': 3,
'mountain': 4,
'plank': 5,
'seatedforwardbend': 6,
'tree': 7,
'trianglepose': 8,
'warrior1': 9,
'warrior2': 10
}
inverse_map = {
1: 'bridge',
2: 'childs',
3: 'downwarddog',
4: 'mountain',
5: 'plank',
6: 'seatedforwardbend',
7: 'tree',
8: 'trianglepose',
9: 'warrior1',
10: 'warrior2'
}
video_file = args.vid_file
# ========= [Optional] download the youtube video ========= #
if video_file.startswith('https://www.youtube.com'):
print(f'Donwloading YouTube video \"{video_file}\"')
video_file = download_youtube_clip(video_file, '/tmp')
if video_file is None:
exit('Youtube url is not valid!')
print(f'YouTube Video has been downloaded to {video_file}...')
if not os.path.isfile(video_file):
exit(f'Input video \"{video_file}\" does not exist!')
dir_path = '/home/ubuntu/PoseEstimation/VIBE/InputData/input_test_set/'
output_folder = '/home/ubuntu/PoseEstimation/VIBE/OutputData/'
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load Classification Model ========= #
classification_model = pickle.load(
open('view_classification_model.pkl', 'rb'))
# ========= Load pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
#print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
#print(f'Loaded pretrained weights from \"{pretrained_file}\"')
image_folder, num_frames, img_shape = video_to_images(video_file,
return_info=True)
print(f'Input video number of frames {num_frames}')
orig_height, orig_width = img_shape[:2]
total_time = time.time()
# ========= Run tracking ========= #
bbox_scale = 1.1
if args.tracking_method == 'pose':
if not os.path.isabs(video_file):
video_file = os.path.join(os.getcwd(), video_file)
tracking_results = run_posetracker(video_file,
staf_folder=args.staf_dir,
display=args.display)
else:
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Run VIBE on each person ========= #
#print(f'Running VIBE on each tracklet...')
vibe_time = time.time()
vibe_results = {}
for person_id in list(tracking_results.keys()):
bboxes = joints2d = None
if args.tracking_method == 'bbox':
bboxes = tracking_results[person_id]['bbox']
elif args.tracking_method == 'pose':
joints2d = tracking_results[person_id]['joints2d']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
dataloader = DataLoader(dataset,
batch_size=args.vibe_batch_size,
num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= [Optional] run Temporal SMPLify to refine the results ========= #
if args.run_smplify and args.tracking_method == 'pose':
norm_joints2d = np.concatenate(norm_joints2d, axis=0)
norm_joints2d = convert_kps(norm_joints2d, src='staf', dst='spin')
norm_joints2d = torch.from_numpy(norm_joints2d).float().to(device)
# Run Temporal SMPLify
update, new_opt_vertices, new_opt_cam, new_opt_pose, new_opt_betas, \
new_opt_joints3d, new_opt_joint_loss, opt_joint_loss = smplify_runner(
pred_rotmat=pred_pose,
pred_betas=pred_betas,
pred_cam=pred_cam,
j2d=norm_joints2d,
device=device,
batch_size=norm_joints2d.shape[0],
pose2aa=False,
)
# update the parameters after refinement
print(
f'Update ratio after Temporal SMPLify: {update.sum()} / {norm_joints2d.shape[0]}'
)
pred_verts = pred_verts.cpu()
pred_cam = pred_cam.cpu()
pred_pose = pred_pose.cpu()
pred_betas = pred_betas.cpu()
pred_joints3d = pred_joints3d.cpu()
pred_verts[update] = new_opt_vertices[update]
pred_cam[update] = new_opt_cam[update]
pred_pose[update] = new_opt_pose[update]
pred_betas[update] = new_opt_betas[update]
pred_joints3d[update] = new_opt_joints3d[update]
elif args.run_smplify and args.tracking_method == 'bbox':
print(
'[WARNING] You need to enable pose tracking to run Temporal SMPLify algorithm!'
)
print('[WARNING] Continuing without running Temporal SMPLify!..')
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
# Runs 1 Euro Filter to smooth out the results
if args.smooth:
min_cutoff = args.smooth_min_cutoff # 0.004
beta = args.smooth_beta # 1.5
print(
f'Running smoothing on person {person_id}, min_cutoff: {min_cutoff}, beta: {beta}'
)
pred_verts, pred_pose, pred_joints3d = smooth_pose(
pred_pose, pred_betas, min_cutoff=min_cutoff, beta=beta)
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
# ========= Extract 3D joint feature for each frame ========= #
list_val = []
for i in range(len(output_dict['joints3d'])):
list_val.append(output_dict['joints3d'][i].flatten().reshape(
1, -1))
input_df = pd.DataFrame(np.concatenate(list_val))
input_df = input_df.round(2)
predicted_classes = classification_model.predict_classes(input_df)
output_df = pd.DataFrame(predicted_classes)
# ========= Printing all possible poses detected for the video ========= #
total_frames = len(output_df)
print(
'\nPrinting probabilities for yoga poses predicted in different frames.'
)
for i, v in output_df.value_counts().items():
val = round((v / total_frames) * 100, 2)
print('Probability of the yoga pose being ' +
inverse_map[i[0]].capitalize() + " is: " + str(val))
print('\nThe yoga pose in the given video is: ' +
inverse_map[output_df[0].value_counts().idxmax()].capitalize())
def main(args):
torch.cuda.set_device(args.gpu_id)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
print(f'Loading video list {args.video_list}')
video_list = [l.strip() for l in open(args.video_list, 'r').readlines()]
if len(video_list) < 1:
print('No files were found in video list')
return
print('Loading VIBE model')
# ========= Define VIBE model ========= #
model = VIBE_Demo(
seqlen=16,
n_layers=2,
hidden_size=1024,
add_linear=True,
use_residual=True,
).to(device)
# ========= Load VIBE pretrained weights ========= #
pretrained_file = download_ckpt(use_3dpw=False)
ckpt = torch.load(pretrained_file)
print(f'Performance of pretrained model on 3DPW: {ckpt["performance"]}')
ckpt = ckpt['gen_state_dict']
model.load_state_dict(ckpt, strict=False)
model.eval()
print(f'Loaded pretrained weights from \"{pretrained_file}\"')
num_videos = len(video_list)
print(f'Processing {num_videos} videos.')
for video_idx, video_file in enumerate(video_list, start=1):
if not osp.isfile(video_file):
print(
f'Input video \"{video_file}\" does not exist! Moving on to next file.'
)
continue
filename = osp.splitext(osp.basename(video_file))[0]
output_path = osp.join(args.output_folder, filename)
os.makedirs(output_path, exist_ok=True)
image_folder, num_frames, img_shape = video_to_images(video_file,
return_info=True)
print(f'[{video_idx}/{num_videos}] Processing {num_frames} frames')
orig_height, orig_width = img_shape[:2]
# ========= Run tracking ========= #
bbox_scale = 1.1
if args.tracking_method == 'pose':
if not osp.isabs(video_file):
video_file = osp.join(os.getcwd(), video_file)
tracking_results = run_posetracker(video_file,
staf_folder=args.staf_dir,
display=args.display)
else:
# run multi object tracker
mot = MPT(
device=device,
batch_size=args.tracker_batch_size,
display=args.display,
detector_type=args.detector,
output_format='dict',
yolo_img_size=args.yolo_img_size,
)
tracking_results = mot(image_folder)
# remove tracklets if num_frames is less than MIN_NUM_FRAMES
for person_id in list(tracking_results.keys()):
if tracking_results[person_id]['frames'].shape[0] < MIN_NUM_FRAMES:
del tracking_results[person_id]
# ========= Run VIBE on each person ========= #
print(f'Running VIBE on each tracklet...')
vibe_results = {}
for person_id in tqdm(list(tracking_results.keys())):
bboxes = joints2d = None
if args.tracking_method == 'bbox':
bboxes = tracking_results[person_id]['bbox']
elif args.tracking_method == 'pose':
joints2d = tracking_results[person_id]['joints2d']
frames = tracking_results[person_id]['frames']
dataset = Inference(
image_folder=image_folder,
frames=frames,
bboxes=bboxes,
joints2d=joints2d,
scale=bbox_scale,
)
bboxes = dataset.bboxes
frames = dataset.frames
has_keypoints = True if joints2d is not None else False
dataloader = DataLoader(dataset,
batch_size=args.vibe_batch_size,
num_workers=16)
with torch.no_grad():
pred_cam, pred_verts, pred_pose, pred_betas, pred_joints3d, norm_joints2d = [], [], [], [], [], []
for batch in dataloader:
if has_keypoints:
batch, nj2d = batch
norm_joints2d.append(nj2d.numpy().reshape(-1, 21, 3))
batch = batch.unsqueeze(0)
batch = batch.to(device)
batch_size, seqlen = batch.shape[:2]
output = model(batch)[-1]
pred_cam.append(output['theta'][:, :, :3].reshape(
batch_size * seqlen, -1))
pred_verts.append(output['verts'].reshape(
batch_size * seqlen, -1, 3))
pred_pose.append(output['theta'][:, :, 3:75].reshape(
batch_size * seqlen, -1))
pred_betas.append(output['theta'][:, :, 75:].reshape(
batch_size * seqlen, -1))
pred_joints3d.append(output['kp_3d'].reshape(
batch_size * seqlen, -1, 3))
pred_cam = torch.cat(pred_cam, dim=0)
pred_verts = torch.cat(pred_verts, dim=0)
pred_pose = torch.cat(pred_pose, dim=0)
pred_betas = torch.cat(pred_betas, dim=0)
pred_joints3d = torch.cat(pred_joints3d, dim=0)
del batch
# ========= [Optional] run Temporal SMPLify to refine the results ========= #
if args.run_smplify and args.tracking_method == 'pose':
norm_joints2d = np.concatenate(norm_joints2d, axis=0)
norm_joints2d = convert_kps(norm_joints2d,
src='staf',
dst='spin')
norm_joints2d = torch.from_numpy(norm_joints2d).float().to(
device)
# Run Temporal SMPLify
update, new_opt_vertices, new_opt_cam, new_opt_pose, new_opt_betas, \
new_opt_joints3d, new_opt_joint_loss, opt_joint_loss = smplify_runner(
pred_rotmat=pred_pose,
pred_betas=pred_betas,
pred_cam=pred_cam,
j2d=norm_joints2d,
device=device,
batch_size=norm_joints2d.shape[0],
pose2aa=False,
)
# update the parameters after refinement
print(
f'Update ratio after Temporal SMPLify: {update.sum()} / {norm_joints2d.shape[0]}'
)
pred_verts = pred_verts.cpu()
pred_cam = pred_cam.cpu()
pred_pose = pred_pose.cpu()
pred_betas = pred_betas.cpu()
pred_joints3d = pred_joints3d.cpu()
pred_verts[update] = new_opt_vertices[update]
pred_cam[update] = new_opt_cam[update]
pred_pose[update] = new_opt_pose[update]
pred_betas[update] = new_opt_betas[update]
pred_joints3d[update] = new_opt_joints3d[update]
elif args.run_smplify and args.tracking_method == 'bbox':
print(
'[WARNING] You need to enable pose tracking to run Temporal SMPLify algorithm!'
)
print(
'[WARNING] Continuing without running Temporal SMPLify!..')
# ========= Save results to a pickle file ========= #
pred_cam = pred_cam.cpu().numpy()
pred_verts = pred_verts.cpu().numpy()
pred_pose = pred_pose.cpu().numpy()
pred_betas = pred_betas.cpu().numpy()
pred_joints3d = pred_joints3d.cpu().numpy()
orig_cam = convert_crop_cam_to_orig_img(cam=pred_cam,
bbox=bboxes,
img_width=orig_width,
img_height=orig_height)
output_dict = {
'pred_cam': pred_cam,
'orig_cam': orig_cam,
'verts': pred_verts,
'pose': pred_pose,
'betas': pred_betas,
'joints3d': pred_joints3d,
'joints2d': joints2d,
'bboxes': bboxes,
'frame_ids': frames,
}
vibe_results[person_id] = output_dict
# Clean-up the temporal folder
shutil.rmtree(image_folder)
# Save the outputs to joblib pkl file. File is loaded through joblib.load(pkl_path)
output_pkl_path = osp.join(args.output_folder, f'{filename}.pkl')
print(f'Saving output results to \"{output_pkl_path}\".')
joblib.dump(vibe_results, output_pkl_path)
# Clean-up after processing
del model
print('================= END =================')
