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()