prediction = camera_to_world(prediction, R=rot, t=0)
# We don't have the trajectory, but at least we can rebase the height
prediction[:, :, 2] -= np.min(prediction[:, :, 2])
anim_output = {'Reconstruction': prediction}
if ground_truth is not None and not args.viz_no_ground_truth:
anim_output['Ground truth'] = ground_truth
input_keypoints = image_coordinates(input_keypoints[..., :2], w=width_of, h=height_of)
manual_fps = 25
from common.visualization import render_animation
render_animation(input_keypoints, anim_output,
dataset.skeleton(), manual_fps, args.viz_bitrate, cam['azimuth'], args.viz_output,
limit=args.viz_limit, downsample=args.viz_downsample, size=args.viz_size,
input_video_path=args.viz_video, viewport=(cam['res_w'], cam['res_h']),
input_video_skip=args.viz_skip)
else:
print('Evaluating...')
all_actions = {}
all_actions_by_subject = {}
for subject in subjects_test:
if subject not in all_actions_by_subject:
all_actions_by_subject[subject] = {}
for action in dataset[subject].keys():
action_name = action.split(' ')[0]
if action_name not in all_actions:
all_actions[action_name] = []
def main(args):
print('==> Using settings {}'.format(args))
print('==> Loading dataset...')
dataset_path = path.join('data', 'data_3d_' + args.dataset + '.npz')
if args.dataset == 'h36m':
from common.h36m_dataset import Human36mDataset
dataset = Human36mDataset(dataset_path)
else:
raise KeyError('Invalid dataset')
print('==> Preparing data...')
dataset = read_3d_data(dataset)
print('==> Loading 2D detections...')
keypoints = create_2d_data(path.join('data', 'data_2d_' + args.dataset + '.npz'), dataset)
cudnn.benchmark = True
device = torch.device("cuda")
# Create model
print("==> Creating model...")
if args.architecture == 'pose_gtac':
from models.pose_gtac import PoseGTAC
from common.graph_utils import adj_mx_from_skeleton
p_dropout = (None if args.dropout == 0.0 else args.dropout)
model_pos = PoseGTAC(args.hid_dim, p_dropout=p_dropout).to(device)
else:
raise KeyError('Invalid model architecture')
print("==> Total parameters: {:.2f}M".format(sum(p.numel() for p in model_pos.parameters()) / 1000000.0))
# Resume from a checkpoint
ckpt_path = args.evaluate
if path.isfile(ckpt_path):
print("==> Loading checkpoint '{}'".format(ckpt_path))
ckpt = torch.load(ckpt_path)
start_epoch = ckpt['epoch']
error_best = ckpt['error']
model_pos.load_state_dict(ckpt['state_dict'])
print("==> Loaded checkpoint (Epoch: {} | Error: {})".format(start_epoch, error_best))
else:
raise RuntimeError("==> No checkpoint found at '{}'".format(ckpt_path))
print('==> Rendering...')
poses_2d = keypoints[args.viz_subject][args.viz_action]
out_poses_2d = poses_2d[args.viz_camera]
out_actions = [args.viz_camera] * out_poses_2d.shape[0]
poses_3d = dataset[args.viz_subject][args.viz_action]['positions_3d']
assert len(poses_3d) == len(poses_2d), 'Camera count mismatch'
out_poses_3d = poses_3d[args.viz_camera]
ground_truth = dataset[args.viz_subject][args.viz_action]['positions_3d'][args.viz_camera].copy()
input_keypoints = out_poses_2d.copy()
render_loader = DataLoader(PoseGenerator([out_poses_3d], [out_poses_2d], [out_actions]), batch_size=args.batch_size,
shuffle=False, num_workers=args.num_workers, pin_memory=True)
prediction = evaluate(render_loader, model_pos, device, args.architecture)[0]
# Invert camera transformation
cam = dataset.cameras()[args.viz_subject][args.viz_camera]
prediction = camera_to_world(prediction, R=cam['orientation'], t=0)
prediction[:, :, 2] -= np.min(prediction[:, :, 2])
ground_truth = camera_to_world(ground_truth, R=cam['orientation'], t=0)
ground_truth[:, :, 2] -= np.min(ground_truth[:, :, 2])
anim_output = {'Regression': prediction, 'Ground truth': ground_truth}
input_keypoints = image_coordinates(input_keypoints[..., :2], w=cam['res_w'], h=cam['res_h'])
render_animation(input_keypoints, anim_output, dataset.skeleton(), dataset.fps(), args.viz_bitrate, cam['azimuth'],
args.viz_output, limit=args.viz_limit, downsample=args.viz_downsample, size=args.viz_size,
input_video_path=args.viz_video, viewport=(cam['res_w'], cam['res_h']),
input_video_skip=args.viz_skip)
anim_output = {'Reconstruction': prediction}
if ground_truth is not None and not args.viz_no_ground_truth:
anim_output['Ground truth'] = ground_truth
input_keypoints = image_coordinates(input_keypoints[..., :2],
w=cam['res_w'],
h=cam['res_h'])
from common.visualization import render_animation
render_animation(input_keypoints,
keypoints_metadata,
anim_output,
dataset.skeleton(),
dataset.fps(),
args.viz_bitrate,
cam['azimuth'],
args.viz_output,
limit=args.viz_limit,
downsample=args.viz_downsample,
size=args.viz_size,
input_video_path=args.viz_video,
viewport=(cam['res_w'], cam['res_h']),
input_video_skip=args.viz_skip)
else:
print('Evaluating...')
all_actions = {}
all_actions_by_subject = {}
for subject in subjects_test:
if subject not in all_actions_by_subject:
all_actions_by_subject[subject] = {}
def the_main_kaboose(args):
print(args)
try:
# Create checkpoint directory if it does not exist
os.makedirs(args.checkpoint)
except OSError as e:
if e.errno != errno.EEXIST:
raise RuntimeError('Unable to create checkpoint directory:',
args.checkpoint)
print('Loading dataset...')
dataset_path = 'data/data_3d_' + args.dataset + '.npz'
if args.dataset == 'h36m':
from common.h36m_dataset import Human36mDataset
dataset = Human36mDataset(dataset_path)
elif args.dataset.startswith('humaneva'):
from common.humaneva_dataset import HumanEvaDataset
dataset = HumanEvaDataset(dataset_path)
elif args.dataset.startswith('custom'):
from common.custom_dataset import CustomDataset
dataset = CustomDataset('data/data_2d_' + args.dataset + '_' +
args.keypoints + '.npz')
else:
raise KeyError('Invalid dataset')
print('Preparing data...')
for subject in dataset.subjects():
for action in dataset[subject].keys():
anim = dataset[subject][action]
# this only works when training.
if 'positions' in anim:
positions_3d = []
for cam in anim['cameras']:
pos_3d = world_to_camera(anim['positions'],
R=cam['orientation'],
t=cam['translation'])
pos_3d[:,
1:] -= pos_3d[:, :
1] # Remove global offset, but keep trajectory in first position
positions_3d.append(pos_3d)
anim['positions_3d'] = positions_3d
print('Loading 2D detections...')
keypoints = np.load('data/data_2d_' + args.dataset + '_' + args.keypoints +
'.npz',
allow_pickle=True)
keypoints_metadata = keypoints['metadata'].item()
keypoints_symmetry = keypoints_metadata['keypoints_symmetry']
kps_left, kps_right = list(keypoints_symmetry[0]), list(
keypoints_symmetry[1])
joints_left, joints_right = list(dataset.skeleton().joints_left()), list(
dataset.skeleton().joints_right())
keypoints = keypoints['positions_2d'].item()
# THIS IS ABOUT TRAINING. ignore pls.
for subject in dataset.subjects():
assert subject in keypoints, 'Subject {} is missing from the 2D detections dataset'.format(
subject)
for action in dataset[subject].keys():
assert action in keypoints[
subject], 'Action {} of subject {} is missing from the 2D detections dataset'.format(
action, subject)
if 'positions_3d' not in dataset[subject][action]:
continue
for cam_idx in range(len(keypoints[subject][action])):
# We check for >= instead of == because some videos in H3.6M contain extra frames
mocap_length = dataset[subject][action]['positions_3d'][
cam_idx].shape[0]
assert keypoints[subject][action][cam_idx].shape[
0] >= mocap_length
if keypoints[subject][action][cam_idx].shape[0] > mocap_length:
# Shorten sequence
keypoints[subject][action][cam_idx] = keypoints[subject][
action][cam_idx][:mocap_length]
assert len(keypoints[subject][action]) == len(
dataset[subject][action]['positions_3d'])
# normalize camera frame?
for subject in keypoints.keys():
for action in keypoints[subject]:
for cam_idx, kps in enumerate(keypoints[subject][action]):
# Normalize camera frame
cam = dataset.cameras()[subject][cam_idx]
kps[..., :2] = normalize_screen_coordinates(kps[..., :2],
w=cam['res_w'],
h=cam['res_h'])
keypoints[subject][action][cam_idx] = kps
subjects_train = args.subjects_train.split(',')
subjects_semi = [] if not args.subjects_unlabeled else args.subjects_unlabeled.split(
',')
if not args.render:
subjects_test = args.subjects_test.split(',')
else:
subjects_test = [args.viz_subject]
semi_supervised = len(subjects_semi) > 0
if semi_supervised and not dataset.supports_semi_supervised():
raise RuntimeError(
'Semi-supervised training is not implemented for this dataset')
def fetch(subjects, action_filter=None, subset=1, parse_3d_poses=True):
out_poses_3d = []
out_poses_2d = []
out_camera_params = []
for subject in subjects:
print("gonna check actions for subject " + subject)
for subject in subjects:
for action in keypoints[subject].keys():
if action_filter is not None:
found = False
for a in action_filter:
if action.startswith(a):
found = True
break
if not found:
continue
poses_2d = keypoints[subject][action]
for i in range(len(poses_2d)): # Iterate across cameras
out_poses_2d.append(poses_2d[i])
if subject in dataset.cameras():
cams = dataset.cameras()[subject]
assert len(cams) == len(poses_2d), 'Camera count mismatch'
for cam in cams:
if 'intrinsic' in cam:
out_camera_params.append(cam['intrinsic'])
if parse_3d_poses and 'positions_3d' in dataset[subject][
action]:
poses_3d = dataset[subject][action]['positions_3d']
assert len(poses_3d) == len(
poses_2d), 'Camera count mismatch'
for i in range(len(poses_3d)): # Iterate across cameras
out_poses_3d.append(poses_3d[i])
if len(out_camera_params) == 0:
out_camera_params = None
if len(out_poses_3d) == 0:
out_poses_3d = None
stride = args.downsample
if subset < 1:
for i in range(len(out_poses_2d)):
n_frames = int(
round(len(out_poses_2d[i]) // stride * subset) * stride)
start = deterministic_random(
0,
len(out_poses_2d[i]) - n_frames + 1,
str(len(out_poses_2d[i])))
out_poses_2d[i] = out_poses_2d[i][start:start +
n_frames:stride]
if out_poses_3d is not None:
out_poses_3d[i] = out_poses_3d[i][start:start +
n_frames:stride]
elif stride > 1:
# Downsample as requested
for i in range(len(out_poses_2d)):
out_poses_2d[i] = out_poses_2d[i][::stride]
if out_poses_3d is not None:
out_poses_3d[i] = out_poses_3d[i][::stride]
return out_camera_params, out_poses_3d, out_poses_2d
action_filter = None if args.actions == '*' else args.actions.split(',')
if action_filter is not None:
print('Selected actions:', action_filter)
# when you run inference, this returns None, None, and the keypoints array renamed as poses_valid_2d
cameras_valid, poses_valid, poses_valid_2d = fetch(subjects_test,
action_filter)
filter_widths = [int(x) for x in args.architecture.split(',')]
if not args.disable_optimizations and not args.dense and args.stride == 1:
# Use optimized model for single-frame predictions
shape_2 = poses_valid_2d[0].shape[-2]
shape_1 = poses_valid_2d[0].shape[-1]
numJoints = dataset.skeleton().num_joints()
model_pos_train = TemporalModelOptimized1f(shape_2,
shape_1,
numJoints,
filter_widths=filter_widths,
causal=args.causal,
dropout=args.dropout,
channels=args.channels)
else:
# When incompatible settings are detected (stride > 1, dense filters, or disabled optimization) fall back to normal model
model_pos_train = TemporalModel(poses_valid_2d[0].shape[-2],
poses_valid_2d[0].shape[-1],
dataset.skeleton().num_joints(),
filter_widths=filter_widths,
causal=args.causal,
dropout=args.dropout,
channels=args.channels,
dense=args.dense)
model_pos = TemporalModel(poses_valid_2d[0].shape[-2],
poses_valid_2d[0].shape[-1],
dataset.skeleton().num_joints(),
filter_widths=filter_widths,
causal=args.causal,
dropout=args.dropout,
channels=args.channels,
dense=args.dense)
receptive_field = model_pos.receptive_field()
print('INFO: Receptive field: {} frames'.format(receptive_field))
pad = (receptive_field - 1) // 2 # Padding on each side
if args.causal:
print('INFO: Using causal convolutions')
causal_shift = pad
else:
causal_shift = 0
model_params = 0
for parameter in model_pos.parameters():
model_params += parameter.numel()
print('INFO: Trainable parameter count:', model_params)
if torch.cuda.is_available():
model_pos = model_pos.cuda()
model_pos_train = model_pos_train.cuda()
if args.resume or args.evaluate:
chk_filename = os.path.join(
args.checkpoint, args.resume if args.resume else args.evaluate)
print('Loading checkpoint', chk_filename)
checkpoint = torch.load(chk_filename,
map_location=lambda storage, loc: storage)
print('This model was trained for {} epochs'.format(
checkpoint['epoch']))
model_pos_train.load_state_dict(checkpoint['model_pos'])
model_pos.load_state_dict(checkpoint['model_pos'])
if args.evaluate and 'model_traj' in checkpoint:
# Load trajectory model if it contained in the checkpoint (e.g. for inference in the wild)
model_traj = TemporalModel(poses_valid_2d[0].shape[-2],
poses_valid_2d[0].shape[-1],
1,
filter_widths=filter_widths,
causal=args.causal,
dropout=args.dropout,
channels=args.channels,
dense=args.dense)
if torch.cuda.is_available():
model_traj = model_traj.cuda()
model_traj.load_state_dict(checkpoint['model_traj'])
else:
model_traj = None
test_generator = UnchunkedGenerator(cameras_valid,
poses_valid,
poses_valid_2d,
pad=pad,
causal_shift=causal_shift,
augment=False,
kps_left=kps_left,
kps_right=kps_right,
joints_left=joints_left,
joints_right=joints_right)
print('INFO: Testing on {} frames'.format(test_generator.num_frames()))
# Evaluate
def evaluate(eval_generator,
action=None,
return_predictions=False,
use_trajectory_model=False):
epoch_loss_3d_pos = 0
epoch_loss_3d_pos_procrustes = 0
epoch_loss_3d_pos_scale = 0
epoch_loss_3d_vel = 0
with torch.no_grad():
if not use_trajectory_model:
model_pos.eval()
else:
model_traj.eval()
N = 0
for _, batch, batch_2d in eval_generator.next_epoch():
inputs_2d = torch.from_numpy(batch_2d.astype('float32'))
if torch.cuda.is_available():
inputs_2d = inputs_2d.cuda()
# Positional model
if not use_trajectory_model:
predicted_3d_pos = model_pos(inputs_2d)
else:
predicted_3d_pos = model_traj(inputs_2d)
# Test-time augmentation (if enabled)
if eval_generator.augment_enabled():
# Undo flipping and take average with non-flipped version
predicted_3d_pos[1, :, :, 0] *= -1
if not use_trajectory_model:
predicted_3d_pos[1, :, joints_left +
joints_right] = predicted_3d_pos[
1, :, joints_right + joints_left]
predicted_3d_pos = torch.mean(predicted_3d_pos,
dim=0,
keepdim=True)
if return_predictions:
return predicted_3d_pos.squeeze(0).cpu().numpy()
inputs_3d = torch.from_numpy(batch.astype('float32'))
if torch.cuda.is_available():
inputs_3d = inputs_3d.cuda()
inputs_3d[:, :, 0] = 0
if eval_generator.augment_enabled():
inputs_3d = inputs_3d[:1]
error = mpjpe(predicted_3d_pos, inputs_3d)
epoch_loss_3d_pos_scale += inputs_3d.shape[
0] * inputs_3d.shape[1] * n_mpjpe(predicted_3d_pos,
inputs_3d).item()
epoch_loss_3d_pos += inputs_3d.shape[0] * inputs_3d.shape[
1] * error.item()
N += inputs_3d.shape[0] * inputs_3d.shape[1]
inputs = inputs_3d.cpu().numpy().reshape(
-1, inputs_3d.shape[-2], inputs_3d.shape[-1])
predicted_3d_pos = predicted_3d_pos.cpu().numpy().reshape(
-1, inputs_3d.shape[-2], inputs_3d.shape[-1])
epoch_loss_3d_pos_procrustes += inputs_3d.shape[
0] * inputs_3d.shape[1] * p_mpjpe(predicted_3d_pos, inputs)
# Compute velocity error
epoch_loss_3d_vel += inputs_3d.shape[0] * inputs_3d.shape[
1] * mean_velocity_error(predicted_3d_pos, inputs)
if action is None:
print('----------')
else:
print('----' + action + '----')
e1 = (epoch_loss_3d_pos / N) * 1000
e2 = (epoch_loss_3d_pos_procrustes / N) * 1000
e3 = (epoch_loss_3d_pos_scale / N) * 1000
ev = (epoch_loss_3d_vel / N) * 1000
print('Test time augmentation:', eval_generator.augment_enabled())
print('Protocol #1 Error (MPJPE):', e1, 'mm')
print('Protocol #2 Error (P-MPJPE):', e2, 'mm')
print('Protocol #3 Error (N-MPJPE):', e3, 'mm')
print('Velocity Error (MPJVE):', ev, 'mm')
print('----------')
return e1, e2, e3, ev
if args.render:
print('Rendering...')
input_keypoints = keypoints[args.viz_subject][args.viz_action][
args.viz_camera].copy()
ground_truth = None
if args.viz_subject in dataset.subjects(
) and args.viz_action in dataset[args.viz_subject]:
if 'positions_3d' in dataset[args.viz_subject][args.viz_action]:
ground_truth = dataset[args.viz_subject][
args.viz_action]['positions_3d'][args.viz_camera].copy()
if ground_truth is None:
print(
'INFO: this action is unlabeled. Ground truth will not be rendered.'
)
gen = UnchunkedGenerator(None,
None, [input_keypoints],
pad=pad,
causal_shift=causal_shift,
augment=args.test_time_augmentation,
kps_left=kps_left,
kps_right=kps_right,
joints_left=joints_left,
joints_right=joints_right)
prediction = evaluate(gen, return_predictions=True)
if model_traj is not None and ground_truth is None:
prediction_traj = evaluate(gen,
return_predictions=True,
use_trajectory_model=True)
prediction += prediction_traj
if args.viz_export is not None:
print('Exporting joint positions to', args.viz_export)
# Predictions are in camera space
np.save(args.viz_export, prediction)
if args.viz_output is not None:
if ground_truth is not None:
# Reapply trajectory
trajectory = ground_truth[:, :1]
ground_truth[:, 1:] += trajectory
prediction += trajectory
# Invert camera transformation
cam = dataset.cameras()[args.viz_subject][args.viz_camera]
if ground_truth is not None:
prediction = camera_to_world(prediction,
R=cam['orientation'],
t=cam['translation'])
ground_truth = camera_to_world(ground_truth,
R=cam['orientation'],
t=cam['translation'])
else:
# If the ground truth is not available, take the camera extrinsic params from a random subject.
# They are almost the same, and anyway, we only need this for visualization purposes.
for subject in dataset.cameras():
if 'orientation' in dataset.cameras()[subject][
args.viz_camera]:
rot = dataset.cameras()[subject][
args.viz_camera]['orientation']
break
prediction = camera_to_world(prediction, R=rot, t=0)
# We don't have the trajectory, but at least we can rebase the height
prediction[:, :, 2] -= np.min(prediction[:, :, 2])
anim_output = {'Reconstruction': prediction}
if ground_truth is not None and not args.viz_no_ground_truth:
anim_output['Ground truth'] = ground_truth
input_keypoints = image_coordinates(input_keypoints[..., :2],
w=cam['res_w'],
h=cam['res_h'])
print("Writing to json")
import json
# format the data in the same format as mediapipe, so we can load it in unity with the same script
# we need a list (frames) of lists of 3d landmarks.
# but prediction[] only has 17 landmarks, and we need 25 in our unity script
unity_landmarks = prediction.tolist()
with open(args.output_json, "w") as json_file:
json.dump(unity_landmarks, json_file)
if args.rendervideo == "yes":
from common.visualization import render_animation
render_animation(input_keypoints,
keypoints_metadata,
anim_output,
dataset.skeleton(),
dataset.fps(),
args.viz_bitrate,
cam['azimuth'],
args.viz_output,
limit=args.viz_limit,
downsample=args.viz_downsample,
size=args.viz_size,
input_video_path=args.viz_video,
viewport=(cam['res_w'], cam['res_h']),
input_video_skip=args.viz_skip)
def main(args):
assert alpha_pose, 'detector_2d should be in ({alpha, hr, open}_pose)'
# 2D kpts loads or generate
if not args.input_npz:
video_name = args.viz_video
keypoints = alpha_pose(video_name)
else:
npz = np.load(args.input_npz)
keypoints = npz['kpts'] # (N, 17, 2)
keypoints_symmetry = metadata['keypoints_symmetry']
kps_left, kps_right = list(keypoints_symmetry[0]), list(keypoints_symmetry[1])
joints_left, joints_right = list([4, 5, 6, 11, 12, 13]), list([1, 2, 3, 14, 15, 16])
# normlization keypoints Suppose using the camera parameter
keypoints = normalize_screen_coordinates(keypoints[..., :2], w=1000, h=1002)
model_pos = TemporalModel(17, 2, 17, filter_widths=[3, 3, 3, 3, 3], causal=args.causal, dropout=args.dropout, channels=args.channels, dense=args.dense)
if torch.cuda.is_available():
model_pos = model_pos.cuda()
ckpt, time1 = ckpt_time(time0)
print('-------------- load data spends {:.2f} seconds'.format(ckpt))
# load trained model
chk_filename = os.path.join(args.checkpoint, args.resume if args.resume else args.evaluate)
print('Loading checkpoint', chk_filename)
checkpoint = torch.load(chk_filename, map_location=lambda storage, loc: storage) # 把loc映射到storage
model_pos.load_state_dict(checkpoint['model_pos'])
ckpt, time2 = ckpt_time(time1)
print('-------------- load 3D model spends {:.2f} seconds'.format(ckpt))
# Receptive field: 243 frames for args.arc [3, 3, 3, 3, 3]
receptive_field = model_pos.receptive_field()
pad = (receptive_field - 1) // 2 # Padding on each side
causal_shift = 0
print('Rendering...')
input_keypoints = keypoints.copy()
gen = UnchunkedGenerator(None, None, [input_keypoints], pad=pad, causal_shift=causal_shift, augment=args.test_time_augmentation,
kps_left=kps_left, kps_right=kps_right, joints_left=joints_left, joints_right=joints_right)
prediction = evaluate(gen, model_pos, return_predictions=True)
# save 3D joint points
np.save('outputs/{0}_3d_output.npy'.format(args.basename), prediction, allow_pickle=True)
rot = np.array([0.14070565, -0.15007018, -0.7552408, 0.62232804], dtype=np.float32)
prediction = camera_to_world(prediction, R=rot, t=0)
# We don't have the trajectory, but at least we can rebase the height
prediction[:, :, 2] -= np.min(prediction[:, :, 2])
anim_output = {'Reconstruction': prediction}
input_keypoints = image_coordinates(input_keypoints[..., :2], w=1000, h=1002)
ckpt, time3 = ckpt_time(time2)
print('-------------- generate reconstruction 3D data spends {:.2f} seconds'.format(ckpt))
from common.visualization import render_animation
render_animation(input_keypoints, anim_output, Skeleton(), 25, args.viz_bitrate, np.array(70., dtype=np.float32), args.viz_output,
limit=args.viz_limit, downsample=args.viz_downsample, size=args.viz_size,
input_video_path=args.viz_video, viewport=(1000, 1002), input_video_skip=args.viz_skip)
ckpt, time4 = ckpt_time(time3)
print('total spend {:2f} second'.format(ckpt))
ckpt, time3 = ckpt_time(time2)
print('------- generate reconstruction 3D data spends {:.2f} seconds'.format(
ckpt))
if not args.viz_output:
args.viz_output = 'xxxx.gif'
# args.viz_limit = 100
from common.visualization import render_animation
render_animation(input_keypoints,
anim_output,
skeleton(),
25,
args.viz_bitrate,
np.array(70., dtype=np.float32),
args.viz_output,
limit=args.viz_limit,
downsample=args.viz_downsample,
size=args.viz_size,
input_video_path=args.viz_video,
viewport=(1000, 1002),
input_video_skip=args.viz_skip)
ckpt, time4 = ckpt_time(time3)
print('------- generaye video spends {:.2f} seconds'.format(ckpt))
ckpt, _ = ckpt_time(time0)
print(' =================== total spends {:.2f} seconds'.format(ckpt))
pace_net.load_weights('weights_pace_network.bin')
model = PoseNetworkLongTerm(30, dataset.skeleton())
if torch.cuda.is_available():
model.cuda()
model.load_weights(long_term_weights_path) # Load pretrained model
if len(sys.argv) == 1:
for subject in dataset.subjects():
for action in dataset[subject].keys():
if '_d0' not in action or '_m' in action:
continue
print('Showing subject %s, action %s.' % (subject, action))
annotated_spline = pace_net.predict(dataset[subject][action]['spline'])
animation = model.generate_motion(annotated_spline, dataset[subject][action])
render_animation(animation, dataset.skeleton(), dataset.fps(), output='interactive')
else:
# Visualize a particular action
action = sys.argv[1]
if action not in dataset[default_subject].keys():
raise ValueError("The specified animation does not exist")
annotated_spline = pace_net.predict(dataset[default_subject][action]['spline'])
animation = model.generate_motion(annotated_spline, dataset[default_subject][action])
if len(sys.argv) == 2:
output_mode = 'interactive'
else:
plt.switch_backend('agg')
output_mode = sys.argv[2]
render_animation(animation, dataset.skeleton(), dataset.fps(), output=output_mode)
def main():
#cap = cv2.VideoCapture(0)
cap = cv2.VideoCapture('D://data//videos//VID_29551_cam0_crop.mkv')
#parser = argparse.ArgumentParser()
opWrapper = op.WrapperPython()
params = dict()
params["model_folder"] = "D://models//"
opWrapper.configure(params)
opWrapper.start()
if not glfw.init():
return
window = glfw.create_window(w_width, w_height, "My OpenGL window", None,
None)
if not window:
glfw.terminate()
return
glfw.make_context_current(window)
glfw.set_window_size_callback(window, window_resize)
vertex_shader = """
#version 330
in vec3 position;
uniform mat4 view;
uniform mat4 model;
uniform mat4 projection;
void main()
{
gl_Position = projection * view * model * vec4(position, 1.0f);
}
"""
fragment_shader = """
#version 330
out vec4 outColor;
void main()
{
outColor = vec4(1.0f,1.0f,1.0f,1.0f);
}
"""
shader = OpenGL.GL.shaders.compileProgram(
OpenGL.GL.shaders.compileShader(vertex_shader, GL_VERTEX_SHADER),
OpenGL.GL.shaders.compileShader(fragment_shader, GL_FRAGMENT_SHADER))
VBO = glGenBuffers(1)
glBindBuffer(GL_ARRAY_BUFFER, VBO)
glBufferData(GL_ARRAY_BUFFER, 17 * 3 * 4, None, GL_DYNAMIC_DRAW)
EBO = glGenBuffers(1)
glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, EBO)
glBufferData(GL_ELEMENT_ARRAY_BUFFER, 32 * 8, parentsIndices,
GL_STATIC_DRAW)
position = glGetAttribLocation(shader, "position")
glVertexAttribPointer(position, 3, GL_FLOAT, GL_FALSE, 0,
ctypes.c_void_p(0))
glEnableVertexAttribArray(position)
glUseProgram(shader)
view = pyrr.matrix44.create_from_translation(pyrr.Vector3([0.0, 0.0,
-3.0]))
projection = pyrr.matrix44.create_perspective_projection_matrix(
45.0, w_width / w_height, 0.1, 100.0)
model = pyrr.matrix44.create_from_translation(pyrr.Vector3([0.0, 0.0,
0.0]))
view_loc = glGetUniformLocation(shader, "view")
proj_loc = glGetUniformLocation(shader, "projection")
model_loc = glGetUniformLocation(shader, "model")
glUniformMatrix4fv(view_loc, 1, GL_FALSE, view)
glUniformMatrix4fv(proj_loc, 1, GL_FALSE, projection)
glUniformMatrix4fv(model_loc, 1, GL_FALSE, model)
glClearColor(114.0 / 255.0, 144.0 / 255.0, 154.0 / 255.0, 1.0)
glEnable(GL_DEPTH_TEST)
glViewport(0, 0, w_width, w_height)
args = parse_args()
print(args)
try:
# Create checkpoint directory if it does not exist
os.makedirs(args.checkpoint)
except OSError as e:
if e.errno != errno.EEXIST:
raise RuntimeError('Unable to create checkpoint directory:',
args.checkpoint)
print('Loading 2D detections...')
keypoints = np.load('data/data_2d_' + args.keypoints + '.npz')
keypoints = keypoints['positions_2d'].item()
subject = 'S1'
action = 'Directions 1'
width_of = 410
height_of = 374
for cam_idx, kps in enumerate(keypoints[subject][action]):
# Normalize camera frame
# cam = dataset.cameras()[subject][cam_idx]
kps[..., :2] = normalize_screen_coordinates(kps[..., :2],
w=width_of,
h=height_of)
keypoints[subject][action][cam_idx] = kps
subjects_train = args.subjects_train.split(',')
subjects_semi = [] if not args.subjects_unlabeled else args.subjects_unlabeled.split(
',')
subjects_test = args.subjects_test.split(',')
semi_supervised = len(subjects_semi) > 0
if semi_supervised and not dataset.supports_semi_supervised():
raise RuntimeError(
'Semi-supervised training is not implemented for this dataset')
action_filter = None if args.actions == '*' else args.actions.split(',')
if action_filter is not None:
print('Selected actions:', action_filter)
cameras_valid, poses_valid, poses_valid_2d = fetch(subjects_test,
keypoints,
args.downsample,
action_filter)
filter_widths = [int(x) for x in args.architecture.split(',')]
# IF RENDERING TO A VIDEO
if args.viz_output:
model_pos = TemporalModel(poses_valid_2d[0].shape[1],
poses_valid_2d[0].shape[2],
17,
filter_widths=filter_widths,
causal=args.causal,
dropout=args.dropout,
channels=args.channels,
dense=args.dense)
else:
model_pos = TemporalModelOptimized1f(poses_valid_2d[0].shape[1],
poses_valid_2d[0].shape[2],
17,
filter_widths=filter_widths,
causal=args.causal,
dropout=args.dropout,
channels=args.channels)
receptive_field = model_pos.receptive_field()
print('INFO: Receptive field: {} frames'.format(receptive_field))
pad = (receptive_field - 1) // 2 # Padding on each side
if args.causal:
print('INFO: Using causal convolutions')
causal_shift = pad
else:
causal_shift = 0
model_params = 0
for parameter in model_pos.parameters():
model_params += parameter.numel()
print('INFO: Trainable parameter count:', model_params)
if torch.cuda.is_available():
model_pos = model_pos.cuda()
# model_pos_train = model_pos_train.cuda()
if args.resume or args.evaluate:
chk_filename = os.path.join(
args.checkpoint, args.resume if args.resume else args.evaluate)
print('Loading checkpoint', chk_filename)
checkpoint = torch.load(chk_filename,
map_location=lambda storage, loc: storage)
print('This model was trained for {} epochs'.format(
checkpoint['epoch']))
model_pos.load_state_dict(checkpoint['model_pos'])
# IF RENDERING TO A VIDEO
if args.viz_output:
print('Rendering...')
my_action = 'Directions 1'
input_keypoints = keypoints[args.viz_subject][my_action][
args.viz_camera].copy()
gen = UnchunkedGenerator(None,
None, [input_keypoints],
pad=pad,
causal_shift=causal_shift,
augment=args.test_time_augmentation,
kps_left=kps_left,
kps_right=kps_right,
joints_left=joints_left,
joints_right=joints_right)
prediction = evaluate(gen, model_pos, return_predictions=True)
ground_truth = None
# these values taken from a camera in the h36m dataset, would be good to get/determine values rom stereo calibration of the pip cameras
prediction = camera_to_world(
prediction,
R=[0.14070565, -0.15007018, -0.7552408, 0.62232804],
t=[1.841107, 4.9552846, 0.5634454])
# We don't have the trajectory, but at least we can rebase the height
prediction[:, :, 2] -= np.min(prediction[:, :, 2])
anim_output = {'Reconstruction': prediction}
input_keypoints = image_coordinates(input_keypoints[..., :2],
w=width_of,
h=height_of)
manual_fps = 25
np.savez('out_3D_vp3d', anim_output['Reconstruction'])
camAzimuth = 70.0
from common.visualization import render_animation
render_animation(input_keypoints,
anim_output,
manual_fps,
args.viz_bitrate,
camAzimuth,
args.viz_output,
limit=args.viz_limit,
downsample=args.viz_downsample,
size=args.viz_size,
input_video_path=args.viz_video,
viewport=(width_of, height_of),
input_video_skip=args.viz_skip)
# IF RENDERING LIVE
else:
print('Rendering...')
my_action = 'Directions 1'
input_keypoints = keypoints[args.viz_subject][my_action][
args.viz_camera].copy()
gen = UnchunkedGenerator(None,
None, [input_keypoints],
pad=pad,
causal_shift=causal_shift,
augment=args.test_time_augmentation,
kps_left=kps_left,
kps_right=kps_right,
joints_left=joints_left,
joints_right=joints_right)
prediction = evaluateLive(gen,
model_pos,
VBO,
window,
model_loc,
cap,
opWrapper,
return_predictions=True)
glfw.terminate()
cap.release()
cv2.destroyAllWindows()
