def test_interhand3d_demo():
# H36M demo
pose_model = init_pose_model(
'configs/hand/3d_kpt_sview_rgb_img/internet/interhand3d/'
'res50_interhand3d_all_256x256.py',
None,
device='cpu')
image_name = 'tests/data/interhand2.6m/image2017.jpg'
det_result = {
'image_name': image_name,
'bbox': [50, 50, 50, 50], # bbox format is 'xywh'
'camera_param': None,
'keypoints_3d_gt': None
}
det_results = [det_result]
dataset = pose_model.cfg.data['test']['type']
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
pose_results = inference_interhand_3d_model(pose_model,
image_name,
det_results,
dataset=dataset)
for res in pose_results:
res['title'] = 'title'
vis_3d_pose_result(
pose_model,
result=pose_results,
img=det_results[0]['image_name'],
dataset_info=dataset_info,
)
# test special cases
# Empty det results
_ = inference_interhand_3d_model(pose_model,
image_name, [],
dataset=dataset)
if torch.cuda.is_available():
_ = inference_interhand_3d_model(pose_model.cuda(),
image_name,
det_results,
dataset=dataset)
with pytest.raises(NotImplementedError):
_ = inference_interhand_3d_model(pose_model,
image_name,
det_results,
dataset='test')
def test_pose_lifter_demo():
# H36M demo
pose_model = init_pose_model(
'configs/body/3d_kpt_sview_rgb_img/pose_lift/'
'h36m/simplebaseline3d_h36m.py',
None,
device='cpu')
pose_det_result = {
'keypoints': np.zeros((17, 3)),
'bbox': [50, 50, 50, 50],
'track_id': 0,
'image_name': 'tests/data/h36m/S1_Directions_1.54138969_000001.jpg',
}
pose_results_2d = [[pose_det_result]]
dataset = pose_model.cfg.data['test']['type']
_ = inference_pose_lifter_model(
pose_model, pose_results_2d, dataset, with_track_id=False)
pose_lift_results = inference_pose_lifter_model(
pose_model, pose_results_2d, dataset, with_track_id=True)
for res in pose_lift_results:
res['title'] = 'title'
vis_3d_pose_result(
pose_model,
pose_lift_results,
img=pose_lift_results[0]['image_name'],
dataset=dataset)
# test special cases
# Empty 2D results
_ = inference_pose_lifter_model(
pose_model, [[]], dataset, with_track_id=False)
if torch.cuda.is_available():
_ = inference_pose_lifter_model(
pose_model.cuda(), pose_results_2d, dataset, with_track_id=False)
with pytest.raises(NotImplementedError):
_ = inference_pose_lifter_model(
pose_model, pose_results_2d, dataset='test')
def test_pose_lifter_demo():
# H36M demo
pose_model = init_pose_model(
'configs/body/3d_kpt_sview_rgb_img/pose_lift/'
'h36m/simplebaseline3d_h36m.py',
None,
device='cpu')
pose_det_result = {
'keypoints': np.zeros((17, 3)),
'bbox': [50, 50, 50, 50],
'track_id': 0,
'image_name': 'tests/data/h36m/S1_Directions_1.54138969_000001.jpg',
}
pose_results_2d = [[pose_det_result]]
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
pose_results_2d = extract_pose_sequence(pose_results_2d,
frame_idx=0,
causal=False,
seq_len=1,
step=1)
_ = inference_pose_lifter_model(pose_model,
pose_results_2d,
dataset_info=dataset_info,
with_track_id=False)
pose_lift_results = inference_pose_lifter_model(pose_model,
pose_results_2d,
dataset_info=dataset_info,
with_track_id=True)
for res in pose_lift_results:
res['title'] = 'title'
vis_3d_pose_result(pose_model,
pose_lift_results,
img=pose_results_2d[0][0]['image_name'],
dataset_info=dataset_info)
# test special cases
# Empty 2D results
_ = inference_pose_lifter_model(pose_model, [[]],
dataset_info=dataset_info,
with_track_id=False)
if torch.cuda.is_available():
_ = inference_pose_lifter_model(pose_model.cuda(),
pose_results_2d,
dataset_info=dataset_info,
with_track_id=False)
# test videopose3d
pose_model = init_pose_model(
'configs/body/3d_kpt_sview_rgb_vid/video_pose_lift/h36m/'
'videopose3d_h36m_243frames_fullconv_supervised_cpn_ft.py',
None,
device='cpu')
pose_det_result_0 = {
'keypoints': np.ones((17, 3)),
'bbox': [50, 50, 100, 100],
'track_id': 0,
'image_name': 'tests/data/h36m/S1_Directions_1.54138969_000001.jpg',
}
pose_det_result_1 = {
'keypoints': np.ones((17, 3)),
'bbox': [50, 50, 100, 100],
'track_id': 1,
'image_name': 'tests/data/h36m/S5_SittingDown.54138969_002061.jpg',
}
pose_det_result_2 = {
'keypoints': np.ones((17, 3)),
'bbox': [50, 50, 100, 100],
'track_id': 2,
'image_name': 'tests/data/h36m/S7_Greeting.55011271_000396.jpg',
}
pose_results_2d = [[pose_det_result_0], [pose_det_result_1],
[pose_det_result_2]]
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
seq_len = pose_model.cfg.test_data_cfg.seq_len
pose_results_2d_seq = extract_pose_sequence(pose_results_2d,
1,
causal=False,
seq_len=seq_len,
step=1)
pose_lift_results = inference_pose_lifter_model(pose_model,
pose_results_2d_seq,
dataset_info=dataset_info,
with_track_id=True,
image_size=[1000, 1000],
norm_pose_2d=True)
for res in pose_lift_results:
res['title'] = 'title'
vis_3d_pose_result(
pose_model,
pose_lift_results,
img=pose_results_2d[0][0]['image_name'],
dataset_info=dataset_info,
)
def main():
parser = ArgumentParser()
parser.add_argument('pose_lifter_config',
help='Config file for the 2nd stage pose lifter model')
parser.add_argument(
'pose_lifter_checkpoint',
help='Checkpoint file for the 2nd stage pose lifter model')
parser.add_argument('--pose-detector-conifig',
type=str,
default=None,
help='Config file for the 1st stage 2D pose detector')
parser.add_argument(
'--pose-detector-checkpoint',
type=str,
default=None,
help='Checkpoint file for the 1st stage 2D pose detector')
parser.add_argument('--img-root', type=str, default='', help='Image root')
parser.add_argument(
'--json-file',
type=str,
default=None,
help='Json file containing image and bbox inforamtion. Optionally,'
'The Jons file can also contain 2D pose information. See'
'"only-second-stage"')
parser.add_argument(
'--camera-param-file',
type=str,
default=None,
help='Camera parameter file for converting 3D pose predictions from '
' the camera space to to world space. If None, no conversion will be '
'applied.')
parser.add_argument(
'--only-second-stage',
action='store_true',
help='If true, load 2D pose detection result from the Json file and '
'skip the 1st stage. The pose detection model will be ignored.')
parser.add_argument(
'--rebase-keypoint-height',
action='store_true',
help='Rebase the predicted 3D pose so its lowest keypoint has a '
'height of 0 (landing on the ground). This is useful for '
'visualization when the model do not predict the global position '
'of the 3D pose.')
parser.add_argument(
'--show-ground-truth',
action='store_true',
help='If True, show ground truth if it is available. The ground truth '
'should be contained in the annotations in the Json file with the key '
'"keypoints_3d" for each instance.')
parser.add_argument('--show',
action='store_true',
default=False,
help='whether to show img')
parser.add_argument('--out-img-root',
type=str,
default=None,
help='Root of the output visualization images. '
'Default not saving the visualization images.')
parser.add_argument('--device',
default='cuda:0',
help='Device for inference')
parser.add_argument('--kpt-thr', type=float, default=0.3)
parser.add_argument('--radius',
type=int,
default=4,
help='Keypoint radius for visualization')
parser.add_argument('--thickness',
type=int,
default=1,
help='Link thickness for visualization')
args = parser.parse_args()
assert args.show or (args.out_img_root != '')
coco = COCO(args.json_file)
# First stage: 2D pose detection
pose_det_results_list = []
if args.only_second_stage:
from mmpose.apis.inference import _xywh2xyxy
print('Stage 1: load 2D pose results from Json file.')
for image_id, image in coco.imgs.items():
image_name = osp.join(args.img_root, image['file_name'])
ann_ids = coco.getAnnIds(image_id)
pose_det_results = []
for ann_id in ann_ids:
ann = coco.anns[ann_id]
keypoints = np.array(ann['keypoints']).reshape(-1, 3)
keypoints[..., 2] = keypoints[..., 2] >= 1
keypoints_3d = np.array(ann['keypoints_3d']).reshape(-1, 4)
keypoints_3d[..., 3] = keypoints_3d[..., 3] >= 1
bbox = np.array(ann['bbox']).reshape(1, -1)
pose_det_result = {
'image_name': image_name,
'bbox': _xywh2xyxy(bbox),
'keypoints': keypoints,
'keypoints_3d': keypoints_3d
}
pose_det_results.append(pose_det_result)
pose_det_results_list.append(pose_det_results)
else:
print('Stage 1: 2D pose detection.')
pose_det_model = init_pose_model(args.pose_detector_config,
args.pose_detector_checkpoint,
device=args.device.lower())
assert pose_det_model.cfg.model.type == 'TopDown', 'Only "TopDown"' \
'model is supported for the 1st stage (2D pose detection)'
dataset = pose_det_model.cfg.data['test']['type']
img_keys = list(coco.imgs.keys())
for i in mmcv.track_iter_progress(range(len(img_keys))):
# get bounding box annotations
image_id = img_keys[i]
image = coco.loadImgs(image_id)[0]
image_name = osp.join(args.img_root, image['file_name'])
ann_ids = coco.getAnnIds(image_id)
# make person results for single image
person_results = []
for ann_id in ann_ids:
person = {}
ann = coco.anns[ann_id]
person['bbox'] = ann['bbox']
person_results.append(person)
pose_det_results, _ = inference_top_down_pose_model(
pose_det_model,
image_name,
person_results,
bbox_thr=None,
format='xywh',
dataset=dataset,
return_heatmap=False,
outputs=None)
for res in pose_det_results:
res['image_name'] = image_name
pose_det_results_list.append(pose_det_results)
# Second stage: Pose lifting
print('Stage 2: 2D-to-3D pose lifting.')
pose_lift_model = init_pose_model(args.pose_lifter_config,
args.pose_lifter_checkpoint,
device=args.device.lower())
assert pose_lift_model.cfg.model.type == 'PoseLifter', 'Only' \
'"PoseLifter" model is supported for the 2nd stage ' \
'(2D-to-3D lifting)'
dataset = pose_lift_model.cfg.data['test']['type']
camera_params = None
if args.camera_param_file is not None:
camera_params = mmcv.load(args.camera_param_file)
for i, pose_det_results in enumerate(
mmcv.track_iter_progress(pose_det_results_list)):
# 2D-to-3D pose lifting
# Note that the pose_det_results are regarded as a single-frame pose
# sequence
pose_lift_results = inference_pose_lifter_model(
pose_lift_model,
pose_results_2d=[pose_det_results],
dataset=dataset,
with_track_id=False)
image_name = pose_det_results[0]['image_name']
# Pose processing
pose_lift_results_vis = []
for idx, res in enumerate(pose_lift_results):
keypoints_3d = res['keypoints_3d']
# project to world space
if camera_params is not None:
keypoints_3d = _keypoint_camera_to_world(
keypoints_3d,
camera_params=camera_params,
image_name=image_name,
dataset=dataset)
# rebase height (z-axis)
if args.rebase_keypoint_height:
keypoints_3d[..., 2] -= np.min(keypoints_3d[..., 2],
axis=-1,
keepdims=True)
res['keypoints_3d'] = keypoints_3d
# Add title
det_res = pose_det_results[idx]
instance_id = det_res.get('track_id', idx)
res['title'] = f'Prediction ({instance_id})'
pose_lift_results_vis.append(res)
# Add ground truth
if args.show_ground_truth:
if 'keypoints_3d' not in det_res:
print('Fail to show ground truth. Please make sure that'
' the instance annotations from the Json file'
' contain "keypoints_3d".')
else:
gt = res.copy()
gt['keypoints_3d'] = det_res['keypoints_3d']
gt['title'] = f'Ground truth ({instance_id})'
pose_lift_results_vis.append(gt)
# Visualization
if args.out_img_root is None:
out_file = None
else:
os.makedirs(args.out_img_root, exist_ok=True)
out_file = osp.join(args.out_img_root, f'vis_{i}.jpg')
vis_3d_pose_result(pose_lift_model,
result=pose_lift_results_vis,
img=pose_lift_results[0]['image_name'],
out_file=out_file)
def main():
parser = ArgumentParser()
parser.add_argument('det_config', help='Config file for detection')
parser.add_argument('det_checkpoint', help='Checkpoint file for detection')
parser.add_argument(
'pose_detector_config',
type=str,
default=None,
help='Config file for the 1st stage 2D pose detector')
parser.add_argument(
'pose_detector_checkpoint',
type=str,
default=None,
help='Checkpoint file for the 1st stage 2D pose detector')
parser.add_argument(
'pose_lifter_config',
help='Config file for the 2nd stage pose lifter model')
parser.add_argument(
'pose_lifter_checkpoint',
help='Checkpoint file for the 2nd stage pose lifter model')
parser.add_argument(
'--video-path', type=str, default='', help='Video path')
parser.add_argument(
'--rebase-keypoint-height',
action='store_true',
help='Rebase the predicted 3D pose so its lowest keypoint has a '
'height of 0 (landing on the ground). This is useful for '
'visualization when the model do not predict the global position '
'of the 3D pose.')
parser.add_argument(
'--norm-pose-2d',
action='store_true',
help='Scale the bbox (along with the 2D pose) to the average bbox '
'scale of the dataset, and move the bbox (along with the 2D pose) to '
'the average bbox center of the dataset. This is useful when bbox '
'is small, especially in multi-person scenarios.')
parser.add_argument(
'--num-instances',
type=int,
default=-1,
help='The number of 3D poses to be visualized in every frame. If '
'less than 0, it will be set to the number of pose results in the '
'first frame.')
parser.add_argument(
'--show',
action='store_true',
default=False,
help='whether to show visualizations.')
parser.add_argument(
'--out-video-root',
type=str,
default=None,
help='Root of the output video file. '
'Default not saving the visualization video.')
parser.add_argument(
'--device', default='cuda:0', help='Device for inference')
parser.add_argument(
'--det-cat-id',
type=int,
default=1,
help='Category id for bounding box detection model')
parser.add_argument(
'--bbox-thr',
type=float,
default=0.9,
help='Bounding box score threshold')
parser.add_argument('--kpt-thr', type=float, default=0.3)
parser.add_argument(
'--use-oks-tracking', action='store_true', help='Using OKS tracking')
parser.add_argument(
'--tracking-thr', type=float, default=0.3, help='Tracking threshold')
parser.add_argument(
'--euro',
action='store_true',
help='Using One_Euro_Filter for smoothing')
parser.add_argument(
'--radius',
type=int,
default=8,
help='Keypoint radius for visualization')
parser.add_argument(
'--thickness',
type=int,
default=2,
help='Link thickness for visualization')
assert has_mmdet, 'Please install mmdet to run the demo.'
args = parser.parse_args()
assert args.show or (args.out_video_root != '')
assert args.det_config is not None
assert args.det_checkpoint is not None
video = mmcv.VideoReader(args.video_path)
assert video.opened, f'Failed to load video file {args.video_path}'
# First stage: 2D pose detection
print('Stage 1: 2D pose detection.')
person_det_model = init_detector(
args.det_config, args.det_checkpoint, device=args.device.lower())
pose_det_model = init_pose_model(
args.pose_detector_config,
args.pose_detector_checkpoint,
device=args.device.lower())
assert pose_det_model.cfg.model.type == 'TopDown', 'Only "TopDown"' \
'model is supported for the 1st stage (2D pose detection)'
pose_det_dataset = pose_det_model.cfg.data['test']['type']
pose_det_results_list = []
next_id = 0
pose_det_results = []
for frame in video:
pose_det_results_last = pose_det_results
# test a single image, the resulting box is (x1, y1, x2, y2)
mmdet_results = inference_detector(person_det_model, frame)
# keep the person class bounding boxes.
person_det_results = process_mmdet_results(mmdet_results,
args.det_cat_id)
# make person results for single image
pose_det_results, _ = inference_top_down_pose_model(
pose_det_model,
frame,
person_det_results,
bbox_thr=args.bbox_thr,
format='xyxy',
dataset=pose_det_dataset,
return_heatmap=False,
outputs=None)
# get track id for each person instance
pose_det_results, next_id = get_track_id(
pose_det_results,
pose_det_results_last,
next_id,
use_oks=args.use_oks_tracking,
tracking_thr=args.tracking_thr,
use_one_euro=args.euro,
fps=video.fps)
pose_det_results_list.append(copy.deepcopy(pose_det_results))
# Second stage: Pose lifting
print('Stage 2: 2D-to-3D pose lifting.')
pose_lift_model = init_pose_model(
args.pose_lifter_config,
args.pose_lifter_checkpoint,
device=args.device.lower())
assert pose_lift_model.cfg.model.type == 'PoseLifter', \
'Only "PoseLifter" model is supported for the 2nd stage ' \
'(2D-to-3D lifting)'
pose_lift_dataset = pose_lift_model.cfg.data['test']['type']
if args.out_video_root == '':
save_out_video = False
else:
os.makedirs(args.out_video_root, exist_ok=True)
save_out_video = True
if save_out_video:
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
fps = video.fps
writer = None
# convert keypoint definition
for pose_det_results in pose_det_results_list:
for res in pose_det_results:
keypoints = res['keypoints']
res['keypoints'] = covert_keypoint_definition(
keypoints, pose_det_dataset, pose_lift_dataset)
# load temporal padding config from model.data_cfg
if hasattr(pose_lift_model.cfg, 'test_data_cfg'):
data_cfg = pose_lift_model.cfg.test_data_cfg
else:
data_cfg = pose_lift_model.cfg.data_cfg
num_instances = args.num_instances
for i, pose_det_results in enumerate(
mmcv.track_iter_progress(pose_det_results_list)):
# extract and pad input pose2d sequence
pose_results_2d = extract_pose_sequence(
pose_det_results_list,
frame_idx=i,
causal=data_cfg.causal,
seq_len=data_cfg.seq_len,
step=data_cfg.seq_frame_interval)
# 2D-to-3D pose lifting
pose_lift_results = inference_pose_lifter_model(
pose_lift_model,
pose_results_2d=pose_results_2d,
dataset=pose_lift_dataset,
with_track_id=True,
image_size=video.resolution,
norm_pose_2d=args.norm_pose_2d)
# Pose processing
pose_lift_results_vis = []
for idx, res in enumerate(pose_lift_results):
keypoints_3d = res['keypoints_3d']
# exchange y,z-axis, and then reverse the direction of x,z-axis
keypoints_3d = keypoints_3d[..., [0, 2, 1]]
keypoints_3d[..., 0] = -keypoints_3d[..., 0]
keypoints_3d[..., 2] = -keypoints_3d[..., 2]
# rebase height (z-axis)
if args.rebase_keypoint_height:
keypoints_3d[..., 2] -= np.min(
keypoints_3d[..., 2], axis=-1, keepdims=True)
res['keypoints_3d'] = keypoints_3d
# add title
det_res = pose_det_results[idx]
instance_id = det_res['track_id']
res['title'] = f'Prediction ({instance_id})'
# only visualize the target frame
res['keypoints'] = det_res['keypoints']
res['bbox'] = det_res['bbox']
res['track_id'] = instance_id
pose_lift_results_vis.append(res)
# Visualization
if num_instances < 0:
num_instances = len(pose_lift_results_vis)
img_vis = vis_3d_pose_result(
pose_lift_model,
result=pose_lift_results_vis,
img=video[i],
out_file=None,
radius=args.radius,
thickness=args.thickness,
num_instances=num_instances)
if save_out_video:
if writer is None:
writer = cv2.VideoWriter(
osp.join(args.out_video_root,
f'vis_{osp.basename(args.video_path)}'), fourcc,
fps, (img_vis.shape[1], img_vis.shape[0]))
writer.write(img_vis)
if save_out_video:
writer.release()
def main():
parser = ArgumentParser()
parser.add_argument('pose_config', help='Config file for pose network')
parser.add_argument('pose_checkpoint', help='Checkpoint file')
parser.add_argument('--img-root', type=str, default='', help='Image root')
parser.add_argument('--json-file',
type=str,
default='',
help='Json file containing image info.')
parser.add_argument(
'--camera-param-file',
type=str,
default=None,
help='Camera parameter file for converting 3D pose predictions from '
' the pixel space to camera space. If None, keypoints in pixel space'
'will be visualized')
parser.add_argument(
'--gt-joints-file',
type=str,
default=None,
help='Optional arguement. Ground truth 3D keypoint parameter file. '
'If None, gt keypoints will not be shown and keypoints in pixel '
'space will be visualized.')
parser.add_argument(
'--rebase-keypoint-height',
action='store_true',
help='Rebase the predicted 3D pose so its lowest keypoint has a '
'height of 0 (landing on the ground). This is useful for '
'visualization when the model do not predict the global position '
'of the 3D pose.')
parser.add_argument(
'--show-ground-truth',
action='store_true',
help='If True, show ground truth keypoint if it is available.')
parser.add_argument('--show',
action='store_true',
default=False,
help='whether to show img')
parser.add_argument('--out-img-root',
type=str,
default=None,
help='Root of the output visualization images. '
'Default not saving the visualization images.')
parser.add_argument('--device',
default='cuda:0',
help='Device for inference')
parser.add_argument('--kpt-thr',
type=float,
default=0.3,
help='Keypoint score threshold')
parser.add_argument('--radius',
type=int,
default=4,
help='Keypoint radius for visualization')
parser.add_argument('--thickness',
type=int,
default=1,
help='Link thickness for visualization')
args = parser.parse_args()
assert args.show or (args.out_img_root != '')
coco = COCO(args.json_file)
# build the pose model from a config file and a checkpoint file
pose_model = init_pose_model(args.pose_config,
args.pose_checkpoint,
device=args.device.lower())
dataset = pose_model.cfg.data['test']['type']
# load camera parameters
camera_params = None
if args.camera_param_file is not None:
camera_params = mmcv.load(args.camera_param_file)
# load ground truth joints parameters
gt_joint_params = None
if args.gt_joints_file is not None:
gt_joint_params = mmcv.load(args.gt_joints_file)
# load hand bounding boxes
det_results_list = []
for image_id, image in coco.imgs.items():
image_name = osp.join(args.img_root, image['file_name'])
ann_ids = coco.getAnnIds(image_id)
det_results = []
capture_key = str(image['capture'])
camera_key = image['camera']
frame_idx = image['frame_idx']
for ann_id in ann_ids:
ann = coco.anns[ann_id]
if camera_params is not None:
camera_param = {
key: camera_params[capture_key][key][camera_key]
for key in camera_params[capture_key].keys()
}
camera_param = _transform_interhand_camera_param(camera_param)
else:
camera_param = None
if gt_joint_params is not None:
joint_param = gt_joint_params[capture_key][str(frame_idx)]
gt_joint = np.concatenate([
np.array(joint_param['world_coord']),
np.array(joint_param['joint_valid'])
],
axis=-1)
else:
gt_joint = None
det_result = {
'image_name': image_name,
'bbox': ann['bbox'], # bbox format is 'xywh'
'camera_param': camera_param,
'keypoints_3d_gt': gt_joint
}
det_results.append(det_result)
det_results_list.append(det_results)
for i, det_results in enumerate(
mmcv.track_iter_progress(det_results_list)):
image_name = det_results[0]['image_name']
pose_results = inference_interhand_3d_model(pose_model,
image_name,
det_results,
dataset=dataset)
# Post processing
pose_results_vis = []
for idx, res in enumerate(pose_results):
keypoints_3d = res['keypoints_3d']
# normalize kpt score
if keypoints_3d[:, 3].max() > 1:
keypoints_3d[:, 3] /= 255
# get 2D keypoints in pixel space
res['keypoints'] = keypoints_3d[:, [0, 1, 3]]
# For model-predicted keypoints, channel 0 and 1 are coordinates
# in pixel space, and channel 2 is the depth (in mm) relative
# to root joints.
# If both camera parameter and absolute depth of root joints are
# provided, we can transform keypoint to camera space for better
# visualization.
camera_param = res['camera_param']
keypoints_3d_gt = res['keypoints_3d_gt']
if camera_param is not None and keypoints_3d_gt is not None:
# build camera model
camera = SimpleCamera(camera_param)
# transform gt joints from world space to camera space
keypoints_3d_gt[:, :3] = camera.world_to_camera(
keypoints_3d_gt[:, :3])
# transform relative depth to absolute depth
keypoints_3d[:21, 2] += keypoints_3d_gt[20, 2]
keypoints_3d[21:, 2] += keypoints_3d_gt[41, 2]
# transform keypoints from pixel space to camera space
keypoints_3d[:, :3] = camera.pixel_to_camera(
keypoints_3d[:, :3])
# rotate the keypoint to make z-axis correspondent to height
# for better visualization
vis_R = np.array([[1, 0, 0], [0, 0, -1], [0, 1, 0]])
keypoints_3d[:, :3] = keypoints_3d[:, :3] @ vis_R
if keypoints_3d_gt is not None:
keypoints_3d_gt[:, :3] = keypoints_3d_gt[:, :3] @ vis_R
# rebase height (z-axis)
if args.rebase_keypoint_height:
valid = keypoints_3d[..., 3] > 0
keypoints_3d[..., 2] -= np.min(keypoints_3d[valid, 2],
axis=-1,
keepdims=True)
res['keypoints_3d'] = keypoints_3d
res['keypoints_3d_gt'] = keypoints_3d_gt
# Add title
instance_id = res.get('track_id', idx)
res['title'] = f'Prediction ({instance_id})'
pose_results_vis.append(res)
# Add ground truth
if args.show_ground_truth:
if keypoints_3d_gt is None:
print('Fail to show ground truth. Please make sure that'
' gt-joints-file is provided.')
else:
gt = res.copy()
if args.rebase_keypoint_height:
valid = keypoints_3d_gt[..., 3] > 0
keypoints_3d_gt[...,
2] -= np.min(keypoints_3d_gt[valid, 2],
axis=-1,
keepdims=True)
gt['keypoints_3d'] = keypoints_3d_gt
gt['title'] = f'Ground truth ({instance_id})'
pose_results_vis.append(gt)
# Visualization
if args.out_img_root is None:
out_file = None
else:
os.makedirs(args.out_img_root, exist_ok=True)
out_file = osp.join(args.out_img_root, f'vis_{i}.jpg')
vis_3d_pose_result(
pose_model,
result=pose_results_vis,
img=det_results[0]['image_name'],
out_file=out_file,
dataset=dataset,
show=args.show,
kpt_score_thr=args.kpt_thr,
radius=args.radius,
thickness=args.thickness,
)
def test_inference_without_dataset_info():
# Top down
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/'
'coco/res50_coco_256x192.py',
None,
device='cpu')
if 'dataset_info' in pose_model.cfg:
_ = pose_model.cfg.pop('dataset_info')
image_name = 'tests/data/coco/000000000785.jpg'
person_result = []
person_result.append({'bbox': [50, 50, 50, 100]})
with pytest.warns(DeprecationWarning):
pose_results, _ = inference_top_down_pose_model(pose_model,
image_name,
person_result,
format='xywh')
with pytest.warns(DeprecationWarning):
vis_pose_result(pose_model, image_name, pose_results)
with pytest.raises(NotImplementedError):
with pytest.warns(DeprecationWarning):
pose_results, _ = inference_top_down_pose_model(pose_model,
image_name,
person_result,
format='xywh',
dataset='test')
# Bottom up
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/associative_embedding/'
'coco/res50_coco_512x512.py',
None,
device='cpu')
if 'dataset_info' in pose_model.cfg:
_ = pose_model.cfg.pop('dataset_info')
image_name = 'tests/data/coco/000000000785.jpg'
with pytest.warns(DeprecationWarning):
pose_results, _ = inference_bottom_up_pose_model(
pose_model, image_name)
with pytest.warns(DeprecationWarning):
vis_pose_result(pose_model, image_name, pose_results)
# Top down tracking
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/'
'coco/res50_coco_256x192.py',
None,
device='cpu')
if 'dataset_info' in pose_model.cfg:
_ = pose_model.cfg.pop('dataset_info')
image_name = 'tests/data/coco/000000000785.jpg'
person_result = [{'bbox': [50, 50, 50, 100]}]
with pytest.warns(DeprecationWarning):
pose_results, _ = inference_top_down_pose_model(pose_model,
image_name,
person_result,
format='xywh')
pose_results, _ = get_track_id(pose_results, [], next_id=0)
with pytest.warns(DeprecationWarning):
vis_pose_tracking_result(pose_model, image_name, pose_results)
with pytest.raises(NotImplementedError):
with pytest.warns(DeprecationWarning):
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset='test')
# Bottom up tracking
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/associative_embedding/'
'coco/res50_coco_512x512.py',
None,
device='cpu')
if 'dataset_info' in pose_model.cfg:
_ = pose_model.cfg.pop('dataset_info')
image_name = 'tests/data/coco/000000000785.jpg'
with pytest.warns(DeprecationWarning):
pose_results, _ = inference_bottom_up_pose_model(
pose_model, image_name)
pose_results, next_id = get_track_id(pose_results, [], next_id=0)
with pytest.warns(DeprecationWarning):
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset='BottomUpCocoDataset')
# Pose lifting
pose_model = init_pose_model(
'configs/body/3d_kpt_sview_rgb_img/pose_lift/'
'h36m/simplebaseline3d_h36m.py',
None,
device='cpu')
pose_det_result = {
'keypoints': np.zeros((17, 3)),
'bbox': [50, 50, 50, 50],
'track_id': 0,
'image_name': 'tests/data/h36m/S1_Directions_1.54138969_000001.jpg',
}
if 'dataset_info' in pose_model.cfg:
_ = pose_model.cfg.pop('dataset_info')
pose_results_2d = [[pose_det_result]]
dataset = pose_model.cfg.data['test']['type']
pose_results_2d = extract_pose_sequence(pose_results_2d,
frame_idx=0,
causal=False,
seq_len=1,
step=1)
with pytest.warns(DeprecationWarning):
_ = inference_pose_lifter_model(pose_model,
pose_results_2d,
dataset,
with_track_id=False)
with pytest.warns(DeprecationWarning):
pose_lift_results = inference_pose_lifter_model(pose_model,
pose_results_2d,
dataset,
with_track_id=True)
for res in pose_lift_results:
res['title'] = 'title'
with pytest.warns(DeprecationWarning):
vis_3d_pose_result(pose_model,
pose_lift_results,
img=pose_results_2d[0][0]['image_name'],
dataset=dataset)
with pytest.raises(NotImplementedError):
with pytest.warns(DeprecationWarning):
_ = inference_pose_lifter_model(pose_model,
pose_results_2d,
dataset='test')
def main():
parser = ArgumentParser()
parser.add_argument('det_config', help='Config file for detection')
parser.add_argument('det_checkpoint', help='Checkpoint file for detection')
parser.add_argument('pose_detector_config',
type=str,
default=None,
help='Config file for the 1st stage 2D pose detector')
parser.add_argument(
'pose_detector_checkpoint',
type=str,
default=None,
help='Checkpoint file for the 1st stage 2D pose detector')
parser.add_argument('pose_lifter_config',
help='Config file for the 2nd stage pose lifter model')
parser.add_argument(
'pose_lifter_checkpoint',
help='Checkpoint file for the 2nd stage pose lifter model')
parser.add_argument('--video-path',
type=str,
default='',
help='Video path')
parser.add_argument(
'--rebase-keypoint-height',
action='store_true',
help='Rebase the predicted 3D pose so its lowest keypoint has a '
'height of 0 (landing on the ground). This is useful for '
'visualization when the model do not predict the global position '
'of the 3D pose.')
parser.add_argument(
'--norm-pose-2d',
action='store_true',
help='Scale the bbox (along with the 2D pose) to the average bbox '
'scale of the dataset, and move the bbox (along with the 2D pose) to '
'the average bbox center of the dataset. This is useful when bbox '
'is small, especially in multi-person scenarios.')
parser.add_argument(
'--num-instances',
type=int,
default=-1,
help='The number of 3D poses to be visualized in every frame. If '
'less than 0, it will be set to the number of pose results in the '
'first frame.')
parser.add_argument('--show',
action='store_true',
default=False,
help='whether to show visualizations.')
parser.add_argument('--out-video-root',
type=str,
default='vis_results',
help='Root of the output video file. '
'Default not saving the visualization video.')
parser.add_argument('--device',
default='cuda:0',
help='Device for inference')
parser.add_argument('--det-cat-id',
type=int,
default=1,
help='Category id for bounding box detection model')
parser.add_argument('--bbox-thr',
type=float,
default=0.9,
help='Bounding box score threshold')
parser.add_argument('--kpt-thr', type=float, default=0.3)
parser.add_argument('--use-oks-tracking',
action='store_true',
help='Using OKS tracking')
parser.add_argument('--tracking-thr',
type=float,
default=0.3,
help='Tracking threshold')
parser.add_argument('--radius',
type=int,
default=8,
help='Keypoint radius for visualization')
parser.add_argument('--thickness',
type=int,
default=2,
help='Link thickness for visualization')
parser.add_argument(
'--smooth',
action='store_true',
help='Apply a temporal filter to smooth the 2D pose estimation '
'results. See also --smooth-filter-cfg.')
parser.add_argument(
'--smooth-filter-cfg',
type=str,
default='configs/_base_/filters/one_euro.py',
help='Config file of the filter to smooth the pose estimation '
'results. See also --smooth.')
parser.add_argument(
'--use-multi-frames',
action='store_true',
default=False,
help='whether to use multi frames for inference in the 2D pose'
'detection stage. Default: False.')
parser.add_argument(
'--online',
action='store_true',
default=False,
help='inference mode. If set to True, can not use future frame'
'information when using multi frames for inference in the 2D pose'
'detection stage. Default: False.')
assert has_mmdet, 'Please install mmdet to run the demo.'
args = parser.parse_args()
assert args.show or (args.out_video_root != '')
assert args.det_config is not None
assert args.det_checkpoint is not None
video = mmcv.VideoReader(args.video_path)
assert video.opened, f'Failed to load video file {args.video_path}'
# First stage: 2D pose detection
print('Stage 1: 2D pose detection.')
print('Initializing model...')
person_det_model = init_detector(args.det_config,
args.det_checkpoint,
device=args.device.lower())
pose_det_model = init_pose_model(args.pose_detector_config,
args.pose_detector_checkpoint,
device=args.device.lower())
assert isinstance(pose_det_model, TopDown), 'Only "TopDown"' \
'model is supported for the 1st stage (2D pose detection)'
# frame index offsets for inference, used in multi-frame inference setting
if args.use_multi_frames:
assert 'frame_indices_test' in pose_det_model.cfg.data.test.data_cfg
indices = pose_det_model.cfg.data.test.data_cfg['frame_indices_test']
pose_det_dataset = pose_det_model.cfg.data['test']['type']
# get datasetinfo
dataset_info = pose_det_model.cfg.data['test'].get('dataset_info', None)
if dataset_info is None:
warnings.warn(
'Please set `dataset_info` in the config.'
'Check https://github.com/open-mmlab/mmpose/pull/663 for details.',
DeprecationWarning)
else:
dataset_info = DatasetInfo(dataset_info)
pose_det_results_list = []
next_id = 0
pose_det_results = []
# whether to return heatmap, optional
return_heatmap = False
# return the output of some desired layers,
# e.g. use ('backbone', ) to return backbone feature
output_layer_names = None
print('Running 2D pose detection inference...')
for frame_id, cur_frame in enumerate(mmcv.track_iter_progress(video)):
pose_det_results_last = pose_det_results
# test a single image, the resulting box is (x1, y1, x2, y2)
mmdet_results = inference_detector(person_det_model, cur_frame)
# keep the person class bounding boxes.
person_det_results = process_mmdet_results(mmdet_results,
args.det_cat_id)
if args.use_multi_frames:
frames = collect_multi_frames(video, frame_id, indices,
args.online)
# make person results for current image
pose_det_results, _ = inference_top_down_pose_model(
pose_det_model,
frames if args.use_multi_frames else cur_frame,
person_det_results,
bbox_thr=args.bbox_thr,
format='xyxy',
dataset=pose_det_dataset,
dataset_info=dataset_info,
return_heatmap=return_heatmap,
outputs=output_layer_names)
# get track id for each person instance
pose_det_results, next_id = get_track_id(
pose_det_results,
pose_det_results_last,
next_id,
use_oks=args.use_oks_tracking,
tracking_thr=args.tracking_thr)
pose_det_results_list.append(copy.deepcopy(pose_det_results))
# Second stage: Pose lifting
print('Stage 2: 2D-to-3D pose lifting.')
print('Initializing model...')
pose_lift_model = init_pose_model(args.pose_lifter_config,
args.pose_lifter_checkpoint,
device=args.device.lower())
assert isinstance(pose_lift_model, PoseLifter), \
'Only "PoseLifter" model is supported for the 2nd stage ' \
'(2D-to-3D lifting)'
pose_lift_dataset = pose_lift_model.cfg.data['test']['type']
if args.out_video_root == '':
save_out_video = False
else:
os.makedirs(args.out_video_root, exist_ok=True)
save_out_video = True
if save_out_video:
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
fps = video.fps
writer = None
# convert keypoint definition
for pose_det_results in pose_det_results_list:
for res in pose_det_results:
keypoints = res['keypoints']
res['keypoints'] = convert_keypoint_definition(
keypoints, pose_det_dataset, pose_lift_dataset)
# load temporal padding config from model.data_cfg
if hasattr(pose_lift_model.cfg, 'test_data_cfg'):
data_cfg = pose_lift_model.cfg.test_data_cfg
else:
data_cfg = pose_lift_model.cfg.data_cfg
# build pose smoother for temporal refinement
if args.smooth:
smoother = Smoother(filter_cfg=args.smooth_filter_cfg,
keypoint_key='keypoints',
keypoint_dim=2)
else:
smoother = None
num_instances = args.num_instances
pose_lift_dataset_info = pose_lift_model.cfg.data['test'].get(
'dataset_info', None)
if pose_lift_dataset_info is None:
warnings.warn(
'Please set `dataset_info` in the config.'
'Check https://github.com/open-mmlab/mmpose/pull/663 for details.',
DeprecationWarning)
else:
pose_lift_dataset_info = DatasetInfo(pose_lift_dataset_info)
print('Running 2D-to-3D pose lifting inference...')
for i, pose_det_results in enumerate(
mmcv.track_iter_progress(pose_det_results_list)):
# extract and pad input pose2d sequence
pose_results_2d = extract_pose_sequence(
pose_det_results_list,
frame_idx=i,
causal=data_cfg.causal,
seq_len=data_cfg.seq_len,
step=data_cfg.seq_frame_interval)
# smooth 2d results
if smoother:
pose_results_2d = smoother.smooth(pose_results_2d)
# 2D-to-3D pose lifting
pose_lift_results = inference_pose_lifter_model(
pose_lift_model,
pose_results_2d=pose_results_2d,
dataset=pose_lift_dataset,
dataset_info=pose_lift_dataset_info,
with_track_id=True,
image_size=video.resolution,
norm_pose_2d=args.norm_pose_2d)
# Pose processing
pose_lift_results_vis = []
for idx, res in enumerate(pose_lift_results):
keypoints_3d = res['keypoints_3d']
# exchange y,z-axis, and then reverse the direction of x,z-axis
keypoints_3d = keypoints_3d[..., [0, 2, 1]]
keypoints_3d[..., 0] = -keypoints_3d[..., 0]
keypoints_3d[..., 2] = -keypoints_3d[..., 2]
# rebase height (z-axis)
if args.rebase_keypoint_height:
keypoints_3d[..., 2] -= np.min(keypoints_3d[..., 2],
axis=-1,
keepdims=True)
res['keypoints_3d'] = keypoints_3d
# add title
det_res = pose_det_results[idx]
instance_id = det_res['track_id']
res['title'] = f'Prediction ({instance_id})'
# only visualize the target frame
res['keypoints'] = det_res['keypoints']
res['bbox'] = det_res['bbox']
res['track_id'] = instance_id
pose_lift_results_vis.append(res)
# Visualization
if num_instances < 0:
num_instances = len(pose_lift_results_vis)
img_vis = vis_3d_pose_result(pose_lift_model,
result=pose_lift_results_vis,
img=video[i],
dataset=pose_lift_dataset,
dataset_info=pose_lift_dataset_info,
out_file=None,
radius=args.radius,
thickness=args.thickness,
num_instances=num_instances,
show=args.show)
if save_out_video:
if writer is None:
writer = cv2.VideoWriter(
osp.join(args.out_video_root,
f'vis_{osp.basename(args.video_path)}'), fourcc,
fps, (img_vis.shape[1], img_vis.shape[0]))
writer.write(img_vis)
if save_out_video:
writer.release()
