def test_bottom_up_pose_tracking_demo():
# COCO demo
# build the pose model from a config file and a checkpoint file
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/associative_embedding/'
'coco/res50_coco_512x512.py',
None,
device='cpu')
image_name = 'tests/data/coco/000000000785.jpg'
pose_results, _ = inference_bottom_up_pose_model(pose_model, image_name)
pose_results, next_id = get_track_id(pose_results, [], next_id=0)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset='BottomUpCocoDataset')
pose_results_last = pose_results
# oks
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id=next_id,
use_oks=True)
pose_results_last = pose_results
# one_euro
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id=next_id,
use_one_euro=True)
def inference_pose():
print('Thread "pose" started')
stop_watch = StopWatch(window=10)
while True:
while len(det_result_queue) < 1:
time.sleep(0.001)
with det_result_queue_mutex:
ts_input, frame, t_info, mmdet_results = det_result_queue.popleft()
pose_results_list = []
for model_info, pose_history in zip(pose_model_list,
pose_history_list):
model_name = model_info['name']
pose_model = model_info['model']
cat_ids = model_info['cat_ids']
pose_results_last = pose_history['pose_results_last']
next_id = pose_history['next_id']
with stop_watch.timeit(model_name):
# process mmdet results
det_results = process_mmdet_results(
mmdet_results,
class_names=det_model.CLASSES,
cat_ids=cat_ids)
# inference pose model
dataset_name = pose_model.cfg.data['test']['type']
pose_results, _ = inference_top_down_pose_model(
pose_model,
frame,
det_results,
bbox_thr=args.det_score_thr,
format='xyxy',
dataset=dataset_name)
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id,
use_oks=False,
tracking_thr=0.3,
use_one_euro=True,
fps=None)
pose_results_list.append(pose_results)
# update pose history
pose_history['pose_results_last'] = pose_results
pose_history['next_id'] = next_id
t_info += stop_watch.report_strings()
with pose_result_queue_mutex:
pose_result_queue.append((ts_input, t_info, pose_results_list))
event_inference_done.set()
def test_bottom_up_pose_tracking_demo():
# COCO demo
# build the pose model from a config file and a checkpoint file
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/associative_embedding/'
'coco/res50_coco_512x512.py',
None,
device='cpu')
image_name = 'tests/data/coco/000000000785.jpg'
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
pose_results, _ = inference_bottom_up_pose_model(pose_model,
image_name,
dataset_info=dataset_info)
pose_results, next_id = get_track_id(pose_results, [], next_id=0)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset_info=dataset_info)
pose_results_last = pose_results
# oks
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id=next_id,
use_oks=True)
pose_results_last = pose_results
# one_euro (will be deprecated)
with pytest.deprecated_call():
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id=next_id,
use_one_euro=True)
def process(self, input_msgs):
input_msg = input_msgs['input']
img = input_msg.get_image()
if self.det_countdown == 0:
# get objects by detection model
self.det_countdown = self.det_interval
preds = inference_detector(self.det_model, img)
objects_det = self._post_process_det(preds)
else:
# get object by pose tracking
objects_det = self._get_objects_by_tracking(img.shape)
self.det_countdown -= 1
objects_pose, _ = inference_top_down_pose_model(self.pose_model,
img,
objects_det,
bbox_thr=self.bbox_thr,
format='xyxy')
objects, next_id = get_track_id(objects_pose,
self.track_info.last_objects,
self.track_info.next_id,
use_oks=False,
tracking_thr=0.3)
self.track_info.next_id = next_id
self.track_info.last_objects = objects.copy()
# Pose smoothing
if self.smoother:
objects = self.smoother.smooth(objects)
for obj in objects:
obj['det_model_cfg'] = self.det_model.cfg
obj['pose_model_cfg'] = self.pose_model.cfg
input_msg.update_objects(objects)
return input_msg
def process(self, input_msgs):
input_msg = input_msgs['input']
img = input_msg.get_image()
if self.class_ids:
objects = input_msg.get_objects(
lambda x: x.get('class_id') in self.class_ids)
elif self.labels:
objects = input_msg.get_objects(
lambda x: x.get('label') in self.labels)
else:
objects = input_msg.get_objects()
# Inference pose
objects, _ = inference_top_down_pose_model(
self.model, img, objects, bbox_thr=self.bbox_thr, format='xyxy')
# Pose tracking
objects, next_id = get_track_id(
objects,
self.track_info.last_objects,
self.track_info.next_id,
use_oks=False,
tracking_thr=0.3)
self.track_info.next_id = next_id
# Copy the prediction to avoid track_info being affected by smoothing
self.track_info.last_objects = [obj.copy() for obj in objects]
# Pose smoothing
if self.smoother:
objects = self.smoother.smooth(objects)
for obj in objects:
obj['pose_model_cfg'] = self.model.cfg
input_msg.update_objects(objects)
return input_msg
def test_pose_tracking_demo():
# COCO demo
# build the pose model from a config file and a checkpoint file
pose_model = init_pose_model(
'configs/top_down/resnet/coco/res50_coco_256x192.py',
None,
device='cpu')
image_name = 'tests/data/coco/000000000785.jpg'
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(pose_model,
image_name,
[[50, 50, 50, 100]],
format='xywh')
pose_results, next_id = get_track_id(pose_results, [], next_id=0)
# show the results
vis_pose_tracking_result(pose_model, image_name, pose_results)
pose_results_last = pose_results
# AIC demo
pose_model = init_pose_model(
'configs/top_down/resnet/aic/res50_aic_256x192.py', None, device='cpu')
image_name = 'tests/data/aic/054d9ce9201beffc76e5ff2169d2af2f027002ca.jpg'
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(
pose_model,
image_name, [[50, 50, 50, 100]],
format='xywh',
dataset='TopDownAicDataset')
pose_results, next_id = get_track_id(pose_results, pose_results_last,
next_id)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset='TopDownAicDataset')
# OneHand10K demo
# build the pose model from a config file and a checkpoint file
pose_model = init_pose_model(
'configs/hand/resnet/onehand10k/res50_onehand10k_256x256.py',
None,
device='cpu')
image_name = 'tests/data/onehand10k/9.jpg'
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(
pose_model,
image_name, [[10, 10, 30, 30]],
format='xywh',
dataset='OneHand10KDataset')
pose_results, next_id = get_track_id(pose_results, pose_results_last,
next_id)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset='OneHand10KDataset')
# InterHand2D demo
pose_model = init_pose_model(
'configs/hand/resnet/interhand2d/res50_interhand2d_all_256x256.py',
None,
device='cpu')
image_name = 'tests/data/interhand2d/image2017.jpg'
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(
pose_model,
image_name, [[50, 50, 0, 0]],
format='xywh',
dataset='InterHand2DDataset')
pose_results, next_id = get_track_id(pose_results, [], next_id=0)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset='InterHand2DDataset')
pose_results_last = pose_results
# MPII demo
pose_model = init_pose_model(
'configs/top_down/resnet/mpii/res50_mpii_256x256.py',
None,
device='cpu')
image_name = 'tests/data/mpii/004645041.jpg'
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(
pose_model,
image_name, [[50, 50, 0, 0]],
format='xywh',
dataset='TopDownMpiiDataset')
pose_results, next_id = get_track_id(pose_results, pose_results_last,
next_id)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset='TopDownMpiiDataset')
with pytest.raises(NotImplementedError):
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset='test')
def main():
"""Visualize the demo images.
Using mmdet to detect the human.
"""
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_config', help='Config file for pose')
parser.add_argument('pose_checkpoint', help='Checkpoint file for pose')
parser.add_argument('--video-path', type=str, help='Video path')
parser.add_argument('--show',
action='store_true',
default=False,
help='whether to show visualizations.')
parser.add_argument('--out-video-root',
default='',
help='Root of the output video file. '
'Default not saving the visualization video.')
parser.add_argument('--device',
default='cuda:0',
help='Device used for inference')
parser.add_argument('--bbox-thr',
type=float,
default=0.3,
help='Bounding box score threshold')
parser.add_argument('--kpt-thr',
type=float,
default=0.3,
help='Keypoint score threshold')
parser.add_argument('--iou-thr',
type=float,
default=0.3,
help='IoU score threshold')
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
det_model = init_detector(args.det_config,
args.det_checkpoint,
device=args.device.lower())
# 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']
cap = cv2.VideoCapture(args.video_path)
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:
fps = cap.get(cv2.CAP_PROP_FPS)
size = (int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)))
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
videoWriter = cv2.VideoWriter(
os.path.join(args.out_video_root,
f'vis_{os.path.basename(args.video_path)}'), fourcc,
fps, size)
# optional
return_heatmap = False
# e.g. use ('backbone', ) to return backbone feature
output_layer_names = None
next_id = 0
pose_results = []
while (cap.isOpened()):
pose_results_last = pose_results
flag, img = cap.read()
if not flag:
break
# test a single image, the resulting box is (x1, y1, x2, y2)
mmdet_results = inference_detector(det_model, img)
# keep the person class bounding boxes.
person_results = process_mmdet_results(mmdet_results)
# test a single image, with a list of bboxes.
pose_results, returned_outputs = inference_top_down_pose_model(
pose_model,
img,
person_results,
bbox_thr=args.bbox_thr,
format='xyxy',
dataset=dataset,
return_heatmap=return_heatmap,
outputs=output_layer_names)
# get track id for each person instance
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id,
iou_thr=args.iou_thr)
# show the results
vis_img = vis_pose_tracking_result(pose_model,
img,
pose_results,
dataset=dataset,
kpt_score_thr=args.kpt_thr,
show=False)
if args.show:
cv2.imshow('Image', vis_img)
if save_out_video:
videoWriter.write(vis_img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cap.release()
if save_out_video:
videoWriter.release()
cv2.destroyAllWindows()
def test_top_down_pose_tracking_demo():
# COCO demo
# build the pose model from a config file and a checkpoint file
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/'
'coco/res50_coco_256x192.py',
None,
device='cpu')
image_name = 'tests/data/coco/000000000785.jpg'
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
person_result = [{'bbox': [50, 50, 50, 100]}]
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(pose_model,
image_name,
person_result,
format='xywh',
dataset_info=dataset_info)
pose_results, next_id = get_track_id(pose_results, [], next_id=0)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset_info=dataset_info)
pose_results_last = pose_results
# AIC demo
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/'
'aic/res50_aic_256x192.py',
None,
device='cpu')
image_name = 'tests/data/aic/054d9ce9201beffc76e5ff2169d2af2f027002ca.jpg'
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(pose_model,
image_name,
person_result,
format='xywh',
dataset_info=dataset_info)
pose_results, next_id = get_track_id(pose_results, pose_results_last,
next_id)
for pose_result in pose_results:
del pose_result['bbox']
pose_results, next_id = get_track_id(pose_results, pose_results_last,
next_id)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset_info=dataset_info)
# OneHand10K demo
# build the pose model from a config file and a checkpoint file
pose_model = init_pose_model(
'configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/'
'onehand10k/res50_onehand10k_256x256.py',
None,
device='cpu')
image_name = 'tests/data/onehand10k/9.jpg'
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(
pose_model,
image_name, [{
'bbox': [10, 10, 30, 30]
}],
format='xywh',
dataset_info=dataset_info)
pose_results, next_id = get_track_id(pose_results, pose_results_last,
next_id)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset_info=dataset_info)
# InterHand2D demo
pose_model = init_pose_model(
'configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/'
'interhand2d/res50_interhand2d_all_256x256.py',
None,
device='cpu')
image_name = 'tests/data/interhand2.6m/image2017.jpg'
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(pose_model,
image_name, [{
'bbox': [50, 50, 0, 0]
}],
format='xywh',
dataset_info=dataset_info)
pose_results, next_id = get_track_id(pose_results, [], next_id=0)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset_info=dataset_info)
pose_results_last = pose_results
# MPII demo
pose_model = init_pose_model(
'configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/'
'mpii/res50_mpii_256x256.py',
None,
device='cpu')
image_name = 'tests/data/mpii/004645041.jpg'
dataset_info = DatasetInfo(pose_model.cfg.data['test']['dataset_info'])
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(pose_model,
image_name, [{
'bbox': [50, 50, 0, 0]
}],
format='xywh',
dataset_info=dataset_info)
pose_results, next_id = get_track_id(pose_results, pose_results_last,
next_id)
# show the results
vis_pose_tracking_result(pose_model,
image_name,
pose_results,
dataset_info=dataset_info)
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():
"""Visualize the demo images.
Using mmdet to detect the human.
"""
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_config', help='Config file for pose')
parser.add_argument('pose_checkpoint', help='Checkpoint file for pose')
parser.add_argument('--video-path', type=str, help='Video path')
parser.add_argument('--show',
action='store_true',
default=False,
help='whether to show visualizations.')
parser.add_argument('--out-video-root',
default='',
help='Root of the output video file. '
'Default not saving the visualization video.')
parser.add_argument('--device',
default='cuda:0',
help='Device used 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=0.3,
help='Bounding box score threshold')
parser.add_argument('--kpt-thr',
type=float,
default=0.3,
help='Keypoint score threshold')
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=4,
help='Keypoint radius for visualization')
parser.add_argument('--thickness',
type=int,
default=1,
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
det_model = init_detector(args.det_config,
args.det_checkpoint,
device=args.device.lower())
# 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']
dataset_info = pose_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)
cap = cv2.VideoCapture(args.video_path)
fps = None
assert cap.isOpened(), f'Faild to load video file {args.video_path}'
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:
fps = cap.get(cv2.CAP_PROP_FPS)
size = (int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)))
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
videoWriter = cv2.VideoWriter(
os.path.join(args.out_video_root,
f'vis_{os.path.basename(args.video_path)}'), fourcc,
fps, size)
# optional
return_heatmap = False
# e.g. use ('backbone', ) to return backbone feature
output_layer_names = None
next_id = 0
pose_results = []
while (cap.isOpened()):
pose_results_last = pose_results
flag, img = cap.read()
if not flag:
break
# test a single image, the resulting box is (x1, y1, x2, y2)
mmdet_results = inference_detector(det_model, img)
# keep the person class bounding boxes.
person_results = process_mmdet_results(mmdet_results, args.det_cat_id)
# test a single image, with a list of bboxes.
pose_results, returned_outputs = inference_top_down_pose_model(
pose_model,
img,
person_results,
bbox_thr=args.bbox_thr,
format='xyxy',
dataset=dataset,
dataset_info=dataset_info,
return_heatmap=return_heatmap,
outputs=output_layer_names)
# get track id for each person instance
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id,
use_oks=args.use_oks_tracking,
tracking_thr=args.tracking_thr,
use_one_euro=args.euro,
fps=fps)
# show the results
vis_img = vis_pose_tracking_result(pose_model,
img,
pose_results,
radius=args.radius,
thickness=args.thickness,
dataset=dataset,
dataset_info=dataset_info,
kpt_score_thr=args.kpt_thr,
show=False)
if args.show:
cv2.imshow('Image', vis_img)
if save_out_video:
videoWriter.write(vis_img)
if args.show and cv2.waitKey(1) & 0xFF == ord('q'):
break
cap.release()
if save_out_video:
videoWriter.release()
if args.show:
cv2.destroyAllWindows()
def main():
"""Visualize the demo images."""
parser = ArgumentParser()
parser.add_argument('pose_config', help='Config file for pose')
parser.add_argument('pose_checkpoint', help='Checkpoint file for pose')
parser.add_argument('--video-path', type=str, help='Video path')
parser.add_argument('--show',
action='store_true',
default=False,
help='whether to show visualizations.')
parser.add_argument('--out-video-root',
default='',
help='Root of the output video file. '
'Default not saving the visualization video.')
parser.add_argument('--device',
default='cuda:0',
help='Device used for inference')
parser.add_argument('--kpt-thr',
type=float,
default=0.5,
help='Keypoint score threshold')
parser.add_argument('--pose-nms-thr',
type=float,
default=0.9,
help='OKS threshold for pose NMS')
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=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_video_root != '')
# 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']
assert (dataset == 'BottomUpCocoDataset')
cap = cv2.VideoCapture(args.video_path)
fps = None
assert cap.isOpened(), f'Faild to load video file {args.video_path}'
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:
fps = cap.get(cv2.CAP_PROP_FPS)
size = (int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)))
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
videoWriter = cv2.VideoWriter(
os.path.join(args.out_video_root,
f'vis_{os.path.basename(args.video_path)}'), fourcc,
fps, size)
# optional
return_heatmap = False
# e.g. use ('backbone', ) to return backbone feature
output_layer_names = None
next_id = 0
pose_results = []
while (cap.isOpened()):
flag, img = cap.read()
if not flag:
break
pose_results_last = pose_results
pose_results, returned_outputs = inference_bottom_up_pose_model(
pose_model,
img,
pose_nms_thr=args.pose_nms_thr,
return_heatmap=return_heatmap,
outputs=output_layer_names)
# get track id for each person instance
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id,
use_oks=args.use_oks_tracking,
tracking_thr=args.tracking_thr,
use_one_euro=args.euro,
fps=fps)
# show the results
vis_img = vis_pose_tracking_result(pose_model,
img,
pose_results,
radius=args.radius,
thickness=args.thickness,
dataset=dataset,
kpt_score_thr=args.kpt_thr,
show=False)
if args.show:
cv2.imshow('Image', vis_img)
if save_out_video:
videoWriter.write(vis_img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cap.release()
if save_out_video:
videoWriter.release()
cv2.destroyAllWindows()
def main():
"""Visualize the demo images."""
parser = ArgumentParser()
parser.add_argument('pose_config', help='Config file for pose')
parser.add_argument('pose_checkpoint', help='Checkpoint file for pose')
parser.add_argument('--video-path', type=str, help='Video path')
parser.add_argument('--show',
action='store_true',
default=False,
help='whether to show visualizations.')
parser.add_argument('--out-video-root',
default='',
help='Root of the output video file. '
'Default not saving the visualization video.')
parser.add_argument('--device',
default='cuda:0',
help='Device used for inference')
parser.add_argument('--kpt-thr',
type=float,
default=0.5,
help='Keypoint score threshold')
parser.add_argument('--pose-nms-thr',
type=float,
default=0.9,
help='OKS threshold for pose NMS')
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='(Deprecated, please use --smooth and --smooth-filter-cfg) '
'Using One_Euro_Filter for smoothing.')
parser.add_argument(
'--smooth',
action='store_true',
help='Apply a temporal filter to smooth the 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('--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_video_root != '')
# 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']
dataset_info = pose_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)
assert (dataset == 'BottomUpCocoDataset')
else:
dataset_info = DatasetInfo(dataset_info)
video = mmcv.VideoReader(args.video_path)
assert video.opened, f'Faild to load video file {args.video_path}'
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:
fps = video.fps
size = (video.width, video.height)
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
videoWriter = cv2.VideoWriter(
os.path.join(args.out_video_root,
f'vis_{os.path.basename(args.video_path)}'), fourcc,
fps, size)
# optional
return_heatmap = False
# build pose smoother for temporal refinement
if args.euro:
warnings.warn(
'Argument --euro will be deprecated in the future. '
'Please use --smooth to enable temporal smoothing, and '
'--smooth-filter-cfg to set the filter config.',
DeprecationWarning)
smoother = Smoother(filter_cfg='configs/_base_/filters/one_euro.py',
keypoint_dim=2)
elif args.smooth:
smoother = Smoother(filter_cfg=args.smooth_filter_cfg, keypoint_dim=2)
else:
smoother = None
# e.g. use ('backbone', ) to return backbone feature
output_layer_names = None
next_id = 0
pose_results = []
for cur_frame in mmcv.track_iter_progress(video):
pose_results_last = pose_results
pose_results, _ = inference_bottom_up_pose_model(
pose_model,
cur_frame,
dataset=dataset,
dataset_info=dataset_info,
pose_nms_thr=args.pose_nms_thr,
return_heatmap=return_heatmap,
outputs=output_layer_names)
# get track id for each person instance
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id,
use_oks=args.use_oks_tracking,
tracking_thr=args.tracking_thr,
use_one_euro=args.euro,
fps=fps)
# post-process the pose results with smoother
if smoother:
pose_results = smoother.smooth(pose_results)
# show the results
vis_frame = vis_pose_tracking_result(pose_model,
cur_frame,
pose_results,
radius=args.radius,
thickness=args.thickness,
dataset=dataset,
dataset_info=dataset_info,
kpt_score_thr=args.kpt_thr,
show=False)
if args.show:
cv2.imshow('Image', vis_frame)
if save_out_video:
videoWriter.write(vis_frame)
if args.show and cv2.waitKey(1) & 0xFF == ord('q'):
break
if save_out_video:
videoWriter.release()
if args.show:
cv2.destroyAllWindows()
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():
"""Visualize the demo images.
Using mmdet to detect the human.
"""
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_config', help='Config file for pose')
parser.add_argument('pose_checkpoint', help='Checkpoint file for pose')
parser.add_argument('--video-path', type=str, help='Video path')
parser.add_argument('--show',
action='store_true',
default=False,
help='whether to show visualizations.')
parser.add_argument('--out-video-root',
default='',
help='Root of the output video file. '
'Default not saving the visualization video.')
parser.add_argument('--device',
default='cuda:0',
help='Device used 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=0.3,
help='Bounding box score threshold')
parser.add_argument('--kpt-thr',
type=float,
default=0.3,
help='Keypoint score threshold')
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='(Deprecated, please use --smooth and --smooth-filter-cfg) '
'Using One_Euro_Filter for smoothing.')
parser.add_argument(
'--smooth',
action='store_true',
help='Apply a temporal filter to smooth the 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('--radius',
type=int,
default=4,
help='Keypoint radius for visualization')
parser.add_argument('--thickness',
type=int,
default=1,
help='Link thickness for visualization')
parser.add_argument(
'--use-multi-frames',
action='store_true',
default=False,
help='whether to use multi frames for inference in the pose'
'estimation 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 pose'
'estimation 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
print('Initializing model...')
det_model = init_detector(args.det_config,
args.det_checkpoint,
device=args.device.lower())
# 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']
dataset_info = pose_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)
# read video
video = mmcv.VideoReader(args.video_path)
assert video.opened, f'Faild to load video file {args.video_path}'
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:
fps = video.fps
size = (video.width, video.height)
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
videoWriter = cv2.VideoWriter(
os.path.join(args.out_video_root,
f'vis_{os.path.basename(args.video_path)}'), fourcc,
fps, size)
# frame index offsets for inference, used in multi-frame inference setting
if args.use_multi_frames:
assert 'frame_indices_test' in pose_model.cfg.data.test.data_cfg
indices = pose_model.cfg.data.test.data_cfg['frame_indices_test']
# build pose smoother for temporal refinement
if args.euro:
warnings.warn(
'Argument --euro will be deprecated in the future. '
'Please use --smooth to enable temporal smoothing, and '
'--smooth-filter-cfg to set the filter config.',
DeprecationWarning)
smoother = Smoother(filter_cfg='configs/_base_/filters/one_euro.py',
keypoint_dim=2)
elif args.smooth:
smoother = Smoother(filter_cfg=args.smooth_filter_cfg, keypoint_dim=2)
else:
smoother = None
# 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
next_id = 0
pose_results = []
print('Running inference...')
for frame_id, cur_frame in enumerate(mmcv.track_iter_progress(video)):
pose_results_last = pose_results
# get the detection results of current frame
# the resulting box is (x1, y1, x2, y2)
mmdet_results = inference_detector(det_model, cur_frame)
# keep the person class bounding boxes.
person_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)
# test a single image, with a list of bboxes.
pose_results, _ = inference_top_down_pose_model(
pose_model,
frames if args.use_multi_frames else cur_frame,
person_results,
bbox_thr=args.bbox_thr,
format='xyxy',
dataset=dataset,
dataset_info=dataset_info,
return_heatmap=return_heatmap,
outputs=output_layer_names)
# get track id for each person instance
pose_results, next_id = get_track_id(pose_results,
pose_results_last,
next_id,
use_oks=args.use_oks_tracking,
tracking_thr=args.tracking_thr)
# post-process the pose results with smoother
if smoother:
pose_results = smoother.smooth(pose_results)
# show the results
vis_frame = vis_pose_tracking_result(pose_model,
cur_frame,
pose_results,
radius=args.radius,
thickness=args.thickness,
dataset=dataset,
dataset_info=dataset_info,
kpt_score_thr=args.kpt_thr,
show=False)
if args.show:
cv2.imshow('Frame', vis_frame)
if save_out_video:
videoWriter.write(vis_frame)
if args.show and cv2.waitKey(1) & 0xFF == ord('q'):
break
if save_out_video:
videoWriter.release()
if args.show:
cv2.destroyAllWindows()
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()
