def run_demo(args): cap = open_images_capture(args.input, args.loop) log.info('OpenVINO Runtime') log.info('\tbuild: {}'.format(get_version())) core = Core() log.info('Reading Object Detection model {}'.format(args.model_od)) detector_person = Detector(core, args.model_od, device=args.device, label_class=args.person_label) log.info('The Object Detection model {} is loaded to {}'.format(args.model_od, args.device)) log.info('Reading Human Pose Estimation model {}'.format(args.model_hpe)) single_human_pose_estimator = HumanPoseEstimator(core, args.model_hpe, device=args.device) log.info('The Human Pose Estimation model {} is loaded to {}'.format(args.model_hpe, args.device)) delay = int(cap.get_type() in ('VIDEO', 'CAMERA')) video_writer = cv2.VideoWriter() frames_processed = 0 presenter = monitors.Presenter(args.utilization_monitors, 25) metrics = PerformanceMetrics() start_time = perf_counter() frame = cap.read() if frame is None: raise RuntimeError("Can't read an image from the input") if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), (frame.shape[1], frame.shape[0])): raise RuntimeError("Can't open video writer") while frame is not None: bboxes = detector_person.detect(frame) human_poses = [single_human_pose_estimator.estimate(frame, bbox) for bbox in bboxes] presenter.drawGraphs(frame) colors = [(0, 0, 255), (255, 0, 0), (0, 255, 0), (255, 0, 0), (0, 255, 0), (255, 0, 0), (0, 255, 0), (255, 0, 0), (0, 255, 0), (255, 0, 0), (0, 255, 0), (255, 0, 0), (0, 255, 0), (255, 0, 0), (0, 255, 0), (255, 0, 0), (0, 255, 0)] for pose, bbox in zip(human_poses, bboxes): cv2.rectangle(frame, (bbox[0], bbox[1]), (bbox[0] + bbox[2], bbox[1] + bbox[3]), (255, 0, 0), 2) for id_kpt, kpt in enumerate(pose): cv2.circle(frame, (int(kpt[0]), int(kpt[1])), 3, colors[id_kpt], -1) metrics.update(start_time, frame) frames_processed += 1 if video_writer.isOpened() and (args.output_limit <= 0 or frames_processed <= args.output_limit): video_writer.write(frame) if not args.no_show: cv2.imshow('Human Pose Estimation Demo', frame) key = cv2.waitKey(delay) if key == 27: break presenter.handleKey(key) start_time = perf_counter() frame = cap.read() metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) next_frame_id = 1 next_frame_id_to_show = 0 metrics = PerformanceMetrics() render_metrics = PerformanceMetrics() video_writer = cv2.VideoWriter() plugin_config = get_user_config(args.device, args.num_streams, args.num_threads) model_adapter = OpenvinoAdapter( create_core(), args.model, device=args.device, plugin_config=plugin_config, max_num_requests=args.num_infer_requests, model_parameters={'input_layouts': args.layout}) start_time = perf_counter() frame = cap.read() if frame is None: raise RuntimeError("Can't read an image from the input") config = { 'target_size': args.tsize, 'aspect_ratio': frame.shape[1] / frame.shape[0], 'confidence_threshold': args.prob_threshold, 'padding_mode': 'center' if args.architecture_type == 'higherhrnet' else None, # the 'higherhrnet' and 'ae' specific 'delta': 0.5 if args.architecture_type == 'higherhrnet' else None, # the 'higherhrnet' and 'ae' specific } model = ImageModel.create_model(ARCHITECTURES[args.architecture_type], model_adapter, config) model.log_layers_info() hpe_pipeline = AsyncPipeline(model) hpe_pipeline.submit_data(frame, 0, { 'frame': frame, 'start_time': start_time }) output_transform = OutputTransform(frame.shape[:2], args.output_resolution) if args.output_resolution: output_resolution = output_transform.new_resolution else: output_resolution = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter( args.utilization_monitors, 55, (round(output_resolution[0] / 4), round(output_resolution[1] / 8))) if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), output_resolution): raise RuntimeError("Can't open video writer") while True: if hpe_pipeline.callback_exceptions: raise hpe_pipeline.callback_exceptions[0] # Process all completed requests results = hpe_pipeline.get_result(next_frame_id_to_show) if results: (poses, scores), frame_meta = results frame = frame_meta['frame'] start_time = frame_meta['start_time'] if len(poses) and args.raw_output_message: print_raw_results(poses, scores, next_frame_id_to_show) presenter.drawGraphs(frame) rendering_start_time = perf_counter() frame = draw_poses(frame, poses, args.prob_threshold, output_transform) render_metrics.update(rendering_start_time) metrics.update(start_time, frame) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) next_frame_id_to_show += 1 if not args.no_show: cv2.imshow('Pose estimation results', frame) key = cv2.waitKey(1) ESC_KEY = 27 # Quit. if key in {ord('q'), ord('Q'), ESC_KEY}: break presenter.handleKey(key) continue if hpe_pipeline.is_ready(): # Get new image/frame start_time = perf_counter() frame = cap.read() if frame is None: break # Submit for inference hpe_pipeline.submit_data(frame, next_frame_id, { 'frame': frame, 'start_time': start_time }) next_frame_id += 1 else: # Wait for empty request hpe_pipeline.await_any() hpe_pipeline.await_all() if hpe_pipeline.callback_exceptions: raise hpe_pipeline.callback_exceptions[0] # Process completed requests for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id): results = hpe_pipeline.get_result(next_frame_id_to_show) (poses, scores), frame_meta = results frame = frame_meta['frame'] start_time = frame_meta['start_time'] if len(poses) and args.raw_output_message: print_raw_results(poses, scores, next_frame_id_to_show) presenter.drawGraphs(frame) rendering_start_time = perf_counter() frame = draw_poses(frame, poses, args.prob_threshold, output_transform) render_metrics.update(rendering_start_time) metrics.update(start_time, frame) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) if not args.no_show: cv2.imshow('Pose estimation results', frame) key = cv2.waitKey(1) ESC_KEY = 27 # Quit. if key in {ord('q'), ord('Q'), ESC_KEY}: break presenter.handleKey(key) metrics.log_total() log_latency_per_stage(cap.reader_metrics.get_latency(), hpe_pipeline.preprocess_metrics.get_latency(), hpe_pipeline.inference_metrics.get_latency(), hpe_pipeline.postprocess_metrics.get_latency(), render_metrics.get_latency()) for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) next_frame_id = 1 next_frame_id_to_show = 0 metrics = PerformanceMetrics() render_metrics = PerformanceMetrics() video_writer = cv2.VideoWriter() if args.adapter == 'openvino': plugin_config = get_user_config(args.device, args.num_streams, args.num_threads) model_adapter = OpenvinoAdapter( create_core(), args.model, device=args.device, plugin_config=plugin_config, max_num_requests=args.num_infer_requests) elif args.adapter == 'ovms': model_adapter = OVMSAdapter(args.model) start_time = perf_counter() frame = cap.read() if frame is None: raise RuntimeError("Can't read an image from the input") model = Deblurring(model_adapter, preload=False) model.reshape(frame.shape) model.log_layers_info() pipeline = AsyncPipeline(model) pipeline.submit_data(frame, 0, {'frame': frame, 'start_time': start_time}) presenter = monitors.Presenter( args.utilization_monitors, 55, (round(frame.shape[1] / 4), round(frame.shape[0] / 8))) if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), (2 * frame.shape[1], frame.shape[0])): raise RuntimeError("Can't open video writer") while True: if pipeline.is_ready(): # Get new image/frame start_time = perf_counter() frame = cap.read() if frame is None: break # Submit for inference pipeline.submit_data(frame, next_frame_id, { 'frame': frame, 'start_time': start_time }) next_frame_id += 1 else: # Wait for empty request pipeline.await_any() if pipeline.callback_exceptions: raise pipeline.callback_exceptions[0] # Process all completed requests results = pipeline.get_result(next_frame_id_to_show) if results: result_frame, frame_meta = results input_frame = frame_meta['frame'] start_time = frame_meta['start_time'] rendering_start_time = perf_counter() if input_frame.shape != result_frame.shape: input_frame = cv2.resize( input_frame, (result_frame.shape[1], result_frame.shape[0])) final_image = cv2.hconcat([input_frame, result_frame]) render_metrics.update(rendering_start_time) presenter.drawGraphs(final_image) metrics.update(start_time, final_image) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(final_image) next_frame_id_to_show += 1 if not args.no_show: cv2.imshow('Deblurring Results', final_image) key = cv2.waitKey(1) if key == 27 or key == 'q' or key == 'Q': break presenter.handleKey(key) pipeline.await_all() # Process completed requests for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id): results = pipeline.get_result(next_frame_id_to_show) while results is None: results = pipeline.get_result(next_frame_id_to_show) result_frame, frame_meta = results input_frame = frame_meta['frame'] start_time = frame_meta['start_time'] rendering_start_time = perf_counter() if input_frame.shape != result_frame.shape: input_frame = cv2.resize( input_frame, (result_frame.shape[1], result_frame.shape[0])) final_image = cv2.hconcat([input_frame, result_frame]) render_metrics.update(rendering_start_time) presenter.drawGraphs(final_image) metrics.update(start_time, final_image) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(final_image) if not args.no_show: cv2.imshow('Deblurring Results', final_image) key = cv2.waitKey(1) metrics.log_total() log_latency_per_stage(cap.reader_metrics.get_latency(), pipeline.preprocess_metrics.get_latency(), pipeline.inference_metrics.get_latency(), pipeline.postprocess_metrics.get_latency(), render_metrics.get_latency()) for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() class_map = load_class_map(args.class_map) if class_map is None: raise RuntimeError("Can't read {}".format(args.class_map)) core = load_core() person_detector = PersonDetector(args.detection_model, args.device, core, num_requests=2, output_shape=DETECTOR_OUTPUT_SHAPE) action_recognizer = ActionRecognizer(args.action_model, args.device, core, num_requests=2, img_scale=ACTION_IMAGE_SCALE, num_classes=len(class_map)) person_tracker = Tracker(person_detector, TRACKER_SCORE_THRESHOLD, TRACKER_IOU_THRESHOLD) video_stream = VideoStream(args.input, ACTION_NET_INPUT_FPS, action_recognizer.input_length) video_stream.start() metrics = PerformanceMetrics() visualizer = Visualizer(VISUALIZER_TRG_FPS) visualizer.register_window('Demo') presenter = monitors.Presenter(args.utilization_monitors) samples_library = None if args.samples_dir is not None and os.path.exists(args.samples_dir): visualizer.register_window('Gesture library') visualizer.start() library_queue = visualizer.get_queue('Gesture library') samples_library = VideoLibrary(args.samples_dir, SAMPLES_MAX_WINDOW_SIZE, list(class_map.values()), library_queue, SAMPLES_TRG_FPS) samples_library.start() else: visualizer.start() last_caption = None active_object_id = -1 tracker_labels_map = {} tracker_labels = set() frames_processed = 0 while True: start_time = perf_counter() frame = video_stream.get_live_frame() batch = video_stream.get_batch() if frame is None or batch is None: break if frames_processed == 0: video_writer = cv2.VideoWriter() if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), video_stream.fps(), (frame.shape[1], frame.shape[0])): raise RuntimeError("Can't open video writer") detections, tracker_labels_map = person_tracker.add_frame( frame, len(OBJECT_IDS), tracker_labels_map) if detections is None: active_object_id = -1 last_caption = None if len(detections) == 1: active_object_id = 0 if active_object_id >= 0: cur_det = [det for det in detections if det.id == active_object_id] if len(cur_det) != 1: active_object_id = -1 last_caption = None continue recognizer_result = action_recognizer(batch, cur_det[0].roi.reshape(-1)) if recognizer_result is not None: action_class_id = np.argmax(recognizer_result) action_class_label = \ class_map[action_class_id] if class_map is not None else action_class_id action_class_score = np.max(recognizer_result) if action_class_score > args.action_threshold: last_caption = 'Last gesture: {} '.format( action_class_label) presenter.drawGraphs(frame) if detections is not None: tracker_labels = {det.id for det in detections} for det in detections: roi_color = (0, 255, 0) if active_object_id == det.id else (128, 128, 128) border_width = 2 if active_object_id == det.id else 1 person_roi = det.roi[0] cv2.rectangle(frame, (person_roi[0], person_roi[1]), (person_roi[2], person_roi[3]), roi_color, border_width) cv2.putText(frame, str(det.id), (person_roi[0] + 10, person_roi[1] + 20), cv2.FONT_HERSHEY_SIMPLEX, 0.8, roi_color, 2) if last_caption is not None: cv2.putText(frame, last_caption, (10, frame.shape[0] - 10), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2) metrics.update(start_time, frame) frames_processed += 1 if video_writer.isOpened() and (args.output_limit <= 0 or frames_processed <= args.output_limit): video_writer.write(frame) if args.no_show: continue visualizer.put_queue(frame, 'Demo') key = visualizer.get_key() if key == 27: # esc break elif key == ord(' '): # space active_object_id = -1 last_caption = None elif key == 13: # enter last_caption = None elif key in OBJECT_IDS: # 0-9 local_bbox_id = int(chr(key)) if local_bbox_id in tracker_labels: active_object_id = local_bbox_id else: presenter.handleKey(key) if samples_library is not None: if key == ord('f'): # forward samples_library.next() elif key == ord('b'): # backward samples_library.prev() if samples_library is not None: samples_library.release() visualizer.release() video_stream.release() metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) # Plugin initialization for specified device and load extensions library if specified log.info('OpenVINO Inference Engine') log.info('\tbuild: {}'.format(get_version())) core = Core() # Read IR log.info('Reading Proposal model {}'.format(args.model_pnet)) p_net = core.read_model(args.model_pnet) if len(p_net.inputs) != 1: raise RuntimeError("Pnet supports only single input topologies") if len(p_net.outputs) != 2: raise RuntimeError("Pnet supports two output topologies") log.info('Reading Refine model {}'.format(args.model_rnet)) r_net = core.read_model(args.model_rnet) if len(r_net.inputs) != 1: raise RuntimeError("Rnet supports only single input topologies") if len(r_net.outputs) != 2: raise RuntimeError("Rnet supports two output topologies") log.info('Reading Output model {}'.format(args.model_onet)) o_net = core.read_model(args.model_onet) if len(o_net.inputs) != 1: raise RuntimeError("Onet supports only single input topologies") if len(o_net.outputs) != 3: raise RuntimeError("Onet supports three output topologies") pnet_input_tensor_name = p_net.inputs[0].get_any_name() rnet_input_tensor_name = r_net.inputs[0].get_any_name() onet_input_tensor_name = o_net.inputs[0].get_any_name() for node in p_net.outputs: if node.shape[1] == 2: pnet_cls_name = node.get_any_name() elif node.shape[1] == 4: pnet_roi_name = node.get_any_name() else: raise RuntimeError("Unsupported output layer for Pnet") for node in r_net.outputs: if node.shape[1] == 2: rnet_cls_name = node.get_any_name() elif node.shape[1] == 4: rnet_roi_name = node.get_any_name() else: raise RuntimeError("Unsupported output layer for Rnet") for node in o_net.outputs: if node.shape[1] == 2: onet_cls_name = node.get_any_name() elif node.shape[1] == 4: onet_roi_name = node.get_any_name() elif node.shape[1] == 10: onet_pts_name = node.get_any_name() else: raise RuntimeError("Unsupported output layer for Onet") next_frame_id = 0 metrics = PerformanceMetrics() presenter = None video_writer = cv2.VideoWriter() is_loaded_before = False while True: start_time = perf_counter() origin_image = cap.read() if origin_image is None: if next_frame_id == 0: raise ValueError("Can't read an image from the input") break if next_frame_id == 0: presenter = monitors.Presenter(args.utilization_monitors, 55, (round(origin_image.shape[1] / 4), round(origin_image.shape[0] / 8))) if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), (origin_image.shape[1], origin_image.shape[0])): raise RuntimeError("Can't open video writer") next_frame_id += 1 rgb_image = cv2.cvtColor(origin_image, cv2.COLOR_BGR2RGB) oh, ow, _ = rgb_image.shape scales = utils.calculate_scales(rgb_image) # ************************************* # Pnet stage # ************************************* pnet_res = [] for i, scale in enumerate(scales): hs = int(oh*scale) ws = int(ow*scale) image = preprocess_image(rgb_image, ws, hs) p_net.reshape({pnet_input_tensor_name: PartialShape([1, 3, ws, hs])}) # Change weidth and height of input blob compiled_pnet = core.compile_model(p_net, args.device) infer_request_pnet = compiled_pnet.create_infer_request() if i == 0 and not is_loaded_before: log.info("The Proposal model {} is loaded to {}".format(args.model_pnet, args.device)) infer_request_pnet.infer(inputs={pnet_input_tensor_name: image}) p_res = {name: infer_request_pnet.get_tensor(name).data[:] for name in {pnet_roi_name, pnet_cls_name}} pnet_res.append(p_res) image_num = len(scales) rectangles = [] for i in range(image_num): roi = pnet_res[i][pnet_roi_name] cls = pnet_res[i][pnet_cls_name] _, _, out_h, out_w = cls.shape out_side = max(out_h, out_w) rectangle = utils.detect_face_12net(cls[0][1], roi[0], out_side, 1/scales[i], ow, oh, score_threshold[0], iou_threshold[0]) rectangles.extend(rectangle) rectangles = utils.NMS(rectangles, iou_threshold[1], 'iou') # Rnet stage if len(rectangles) > 0: r_net.reshape({rnet_input_tensor_name: PartialShape([len(rectangles), 3, 24, 24])}) # Change batch size of input blob compiled_rnet = core.compile_model(r_net, args.device) infer_request_rnet = compiled_rnet.create_infer_request() if not is_loaded_before: log.info("The Refine model {} is loaded to {}".format(args.model_rnet, args.device)) rnet_input = [] for rectangle in rectangles: crop_img = rgb_image[int(rectangle[1]):int(rectangle[3]), int(rectangle[0]):int(rectangle[2])] crop_img = preprocess_image(crop_img, 24, 24) rnet_input.extend(crop_img) infer_request_rnet.infer(inputs={rnet_input_tensor_name: rnet_input}) rnet_res = {name: infer_request_rnet.get_tensor(name).data[:] for name in {rnet_roi_name, rnet_cls_name}} roi = rnet_res[rnet_roi_name] cls = rnet_res[rnet_cls_name] rectangles = utils.filter_face_24net(cls, roi, rectangles, ow, oh, score_threshold[1], iou_threshold[2]) # Onet stage if len(rectangles) > 0: o_net.reshape({onet_input_tensor_name: PartialShape([len(rectangles), 3, 48, 48])}) # Change batch size of input blob compiled_onet = core.compile_model(o_net, args.device) infer_request_onet = compiled_onet.create_infer_request() if not is_loaded_before: log.info("The Output model {} is loaded to {}".format(args.model_onet, args.device)) is_loaded_before = True onet_input = [] for rectangle in rectangles: crop_img = rgb_image[int(rectangle[1]):int(rectangle[3]), int(rectangle[0]):int(rectangle[2])] crop_img = preprocess_image(crop_img, 48, 48) onet_input.extend(crop_img) infer_request_onet.infer(inputs={onet_input_tensor_name: onet_input}) onet_res = {name: infer_request_onet.get_tensor(name).data[:] for name in {onet_roi_name, onet_cls_name, onet_pts_name}} roi = onet_res[onet_roi_name] cls = onet_res[onet_cls_name] pts = onet_res[onet_pts_name] rectangles = utils.filter_face_48net(cls, roi, pts, rectangles, ow, oh, score_threshold[2], iou_threshold[3]) # display results for rectangle in rectangles: # Draw detected boxes cv2.putText(origin_image, 'confidence: {:.2f}'.format(rectangle[4]), (int(rectangle[0]), int(rectangle[1])), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0)) cv2.rectangle(origin_image, (int(rectangle[0]), int(rectangle[1])), (int(rectangle[2]), int(rectangle[3])), (255, 0, 0), 1) # Draw landmarks for i in range(5, 15, 2): cv2.circle(origin_image, (int(rectangle[i+0]), int(rectangle[i+1])), 2, (0, 255, 0)) metrics.update(start_time, origin_image) if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id <= args.output_limit): video_writer.write(origin_image) if not args.no_show: cv2.imshow('MTCNN Results', origin_image) key = cv2.waitKey(1) if key in {ord('q'), ord('Q'), 27}: break presenter.handleKey(key) metrics.log_total()
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) frame_processor = FrameProcessor(args) frame_num = 0 metrics = PerformanceMetrics() presenter = None output_transform = None input_crop = None if args.crop_size[0] > 0 and args.crop_size[1] > 0: input_crop = np.array(args.crop_size) elif not (args.crop_size[0] == 0 and args.crop_size[1] == 0): raise ValueError('Both crop height and width should be positive') video_writer = cv2.VideoWriter() while True: start_time = perf_counter() frame = cap.read() if frame is None: if frame_num == 0: raise ValueError("Can't read an image from the input") break if input_crop: frame = center_crop(frame, input_crop) if frame_num == 0: output_transform = OutputTransform(frame.shape[:2], args.output_resolution) if args.output_resolution: output_resolution = output_transform.new_resolution else: output_resolution = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter(args.utilization_monitors, 55, (round(output_resolution[0] / 4), round(output_resolution[1] / 8))) if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), output_resolution): raise RuntimeError("Can't open video writer") detections = frame_processor.process(frame) presenter.drawGraphs(frame) frame = draw_detections(frame, frame_processor, detections, output_transform) metrics.update(start_time, frame) frame_num += 1 if video_writer.isOpened() and (args.output_limit <= 0 or frame_num <= args.output_limit): video_writer.write(frame) if not args.no_show: cv2.imshow('Face recognition demo', frame) key = cv2.waitKey(1) # Quit if key in {ord('q'), ord('Q'), 27}: break presenter.handleKey(key) metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) if args.adapter == 'openvino': plugin_config = get_user_config(args.device, args.num_streams, args.num_threads) model_adapter = OpenvinoAdapter(create_core(), args.model, device=args.device, plugin_config=plugin_config, max_num_requests=args.num_infer_requests, model_parameters={'input_layouts': args.layout}) elif args.adapter == 'ovms': model_adapter = OVMSAdapter(args.model) configuration = { 'confidence_threshold': args.prob_threshold, 'path_to_labels': args.labels, } model = get_model(model_adapter, configuration) model.log_layers_info() pipeline = AsyncPipeline(model) next_frame_id = 0 next_frame_id_to_show = 0 tracker = None if not args.no_track and cap.get_type() in {'VIDEO', 'CAMERA'}: tracker = StaticIOUTracker() visualizer = InstanceSegmentationVisualizer(model.labels, args.show_boxes, args.show_scores) metrics = PerformanceMetrics() render_metrics = PerformanceMetrics() presenter = None output_transform = None video_writer = cv2.VideoWriter() while True: if pipeline.is_ready(): # Get new image/frame start_time = perf_counter() frame = cap.read() if frame is None: if next_frame_id == 0: raise ValueError("Can't read an image from the input") break if next_frame_id == 0: output_transform = OutputTransform(frame.shape[:2], args.output_resolution) if args.output_resolution: output_resolution = output_transform.new_resolution else: output_resolution = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter(args.utilization_monitors, 55, (round(output_resolution[0] / 4), round(output_resolution[1] / 8))) if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), tuple(output_resolution)): raise RuntimeError("Can't open video writer") # Submit for inference pipeline.submit_data(frame, next_frame_id, {'frame': frame, 'start_time': start_time}) next_frame_id += 1 else: # Wait for empty request pipeline.await_any() if pipeline.callback_exceptions: raise pipeline.callback_exceptions[0] # Process all completed requests results = pipeline.get_result(next_frame_id_to_show) if results: (scores, classes, boxes, masks), frame_meta = results frame = frame_meta['frame'] start_time = frame_meta['start_time'] if args.raw_output_message: print_raw_results(boxes, classes, scores, next_frame_id_to_show) rendering_start_time = perf_counter() masks_tracks_ids = tracker(masks, classes) if tracker else None frame = visualizer(frame, boxes, classes, scores, masks, masks_tracks_ids) render_metrics.update(rendering_start_time) presenter.drawGraphs(frame) metrics.update(start_time, frame) if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) next_frame_id_to_show += 1 if not args.no_show: cv2.imshow('Instance Segmentation results', frame) key = cv2.waitKey(1) if key == 27 or key == 'q' or key == 'Q': break presenter.handleKey(key) pipeline.await_all() if pipeline.callback_exceptions: raise pipeline.callback_exceptions[0] # Process completed requests for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id): results = pipeline.get_result(next_frame_id_to_show) (scores, classes, boxes, masks), frame_meta = results frame = frame_meta['frame'] start_time = frame_meta['start_time'] if args.raw_output_message: print_raw_results(boxes, classes, scores, next_frame_id_to_show) rendering_start_time = perf_counter() masks_tracks_ids = tracker(masks, classes) if tracker else None frame = visualizer(frame, boxes, classes, scores, masks, masks_tracks_ids) render_metrics.update(rendering_start_time) presenter.drawGraphs(frame) metrics.update(start_time, frame) if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) if not args.no_show: cv2.imshow('Instance Segmentation results', frame) cv2.waitKey(1) metrics.log_total() log_latency_per_stage(cap.reader_metrics.get_latency(), pipeline.preprocess_metrics.get_latency(), pipeline.inference_metrics.get_latency(), pipeline.postprocess_metrics.get_latency(), render_metrics.get_latency()) for rep in presenter.reportMeans(): log.info(rep)
def run(params, config, capture, detector, reid): win_name = 'Multi camera tracking' frame_number = 0 output_detections = [[] for _ in range(capture.get_num_sources())] key = -1 if config.normalizer_config.enabled: capture.add_transform( NormalizerCLAHE( config.normalizer_config.clip_limit, config.normalizer_config.tile_size, )) tracker = MultiCameraTracker(capture.get_num_sources(), reid, config.sct_config, **vars(config.mct_config), visual_analyze=config.analyzer) thread_body = FramesThreadBody(capture, max_queue_length=len(capture.captures) * 2) frames_thread = Thread(target=thread_body) frames_thread.start() frames_read = False set_output_params = False prev_frames = thread_body.frames_queue.get() detector.run_async(prev_frames, frame_number) metrics = PerformanceMetrics() presenter = monitors.Presenter(params.utilization_monitors, 0) while thread_body.process: if not params.no_show: key = check_pressed_keys(key) if key == 27: break presenter.handleKey(key) start_time = time.perf_counter() try: frames = thread_body.frames_queue.get_nowait() frames_read = True except queue.Empty: frames = None if frames is None: continue all_detections = detector.wait_and_grab() if params.save_detections: update_detections(output_detections, all_detections, frame_number) frame_number += 1 detector.run_async(frames, frame_number) all_masks = [[] for _ in range(len(all_detections))] for i, detections in enumerate(all_detections): all_detections[i] = [det[0] for det in detections] all_masks[i] = [det[2] for det in detections if len(det) == 3] tracker.process(prev_frames, all_detections, all_masks) tracked_objects = tracker.get_tracked_objects() vis = visualize_multicam_detections( prev_frames, tracked_objects, **vars(config.visualization_config)) metrics.update(start_time, vis) presenter.drawGraphs(vis) if not params.no_show: cv.imshow(win_name, vis) if frames_read and not set_output_params: set_output_params = True if len(params.output_video): frame_size = [frame.shape[::-1] for frame in frames] fps = capture.get_fps() target_width, target_height = get_target_size( frame_size, None, **vars(config.visualization_config)) video_output_size = (target_width, target_height) fourcc = cv.VideoWriter_fourcc(*'XVID') output_video = cv.VideoWriter(params.output_video, fourcc, min(fps), video_output_size) else: output_video = None if set_output_params and output_video: output_video.write(cv.resize(vis, video_output_size)) prev_frames, frames = frames, prev_frames metrics.log_total() for rep in presenter.reportMeans(): log.info(rep) thread_body.process = False frames_thread.join() if len(params.history_file): save_json_file(params.history_file, tracker.get_all_tracks_history(), description='History file') if len(params.save_detections): save_json_file(params.save_detections, output_detections, description='Detections') if len(config.embeddings.save_path): save_embeddings(tracker.scts, **vars(config.embeddings))
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) if args.adapter == 'openvino': plugin_config = get_user_config(args.device, args.num_streams, args.num_threads) model_adapter = OpenvinoAdapter(create_core(), args.model, device=args.device, plugin_config=plugin_config, max_num_requests=args.num_infer_requests) elif args.adapter == 'ovms': model_adapter = OVMSAdapter(args.model) model = MonoDepthModel(model_adapter) model.log_layers_info() pipeline = AsyncPipeline(model) next_frame_id = 0 next_frame_id_to_show = 0 metrics = PerformanceMetrics() presenter = None output_transform = None video_writer = cv2.VideoWriter() while True: if pipeline.is_ready(): # Get new image/frame start_time = perf_counter() frame = cap.read() if frame is None: if next_frame_id == 0: raise ValueError("Can't read an image from the input") break if next_frame_id == 0: output_transform = OutputTransform(frame.shape[:2], args.output_resolution) if args.output_resolution: output_resolution = output_transform.new_resolution else: output_resolution = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter(args.utilization_monitors, 55, (round(output_resolution[0] / 4), round(output_resolution[1] / 8))) if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), output_resolution): raise RuntimeError("Can't open video writer") # Submit for inference pipeline.submit_data(frame, next_frame_id, {'start_time': start_time}) next_frame_id += 1 else: # Wait for empty request pipeline.await_any() if pipeline.callback_exceptions: raise pipeline.callback_exceptions[0] # Process all completed requests results = pipeline.get_result(next_frame_id_to_show) if results: depth_map, frame_meta = results depth_map = apply_color_map(depth_map, output_transform) start_time = frame_meta['start_time'] presenter.drawGraphs(depth_map) metrics.update(start_time, depth_map) if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit-1): video_writer.write(depth_map) next_frame_id_to_show += 1 if not args.no_show: cv2.imshow(DEMO_NAME, depth_map) key = cv2.waitKey(1) if key == 27 or key == 'q' or key == 'Q': break presenter.handleKey(key) pipeline.await_all() # Process completed requests for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id): results = pipeline.get_result(next_frame_id_to_show) while results is None: results = pipeline.get_result(next_frame_id_to_show) depth_map, frame_meta = results depth_map = apply_color_map(depth_map, output_transform) start_time = frame_meta['start_time'] presenter.drawGraphs(depth_map) metrics.update(start_time, depth_map) if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit-1): video_writer.write(depth_map) if not args.no_show: cv2.imshow(DEMO_NAME, depth_map) key = cv2.waitKey(1) if key == 27 or key == 'q' or key == 'Q': break presenter.handleKey(key) metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)
def main(): parser = argparse.ArgumentParser(description='Whiteboard inpainting demo') parser.add_argument('-i', '--input', required=True, help='Required. Path to a video file or a device node of a web-camera.') parser.add_argument('--loop', default=False, action='store_true', help='Optional. Enable reading the input in a loop.') parser.add_argument('-o', '--output', required=False, help='Optional. Name of the output file(s) to save.') parser.add_argument('-limit', '--output_limit', required=False, default=1000, type=int, help='Optional. Number of frames to store in output. ' 'If 0 is set, all frames are stored.') parser.add_argument('-m_i', '--m_instance_segmentation', type=str, required=False, help='Required. Path to the instance segmentation model.') parser.add_argument('-m_s', '--m_semantic_segmentation', type=str, required=False, help='Required. Path to the semantic segmentation model.') parser.add_argument('-t', '--threshold', type=float, default=0.6, help='Optional. Threshold for person instance segmentation model.') parser.add_argument('--no_show', help="Optional. Don't show output.", action='store_true') parser.add_argument('-d', '--device', type=str, default='CPU', help='Optional. Specify a target device to infer on. CPU, GPU, HDDL or MYRIAD is ' 'acceptable. The demo will look for a suitable plugin for the device specified.') parser.add_argument('-l', '--cpu_extension', type=str, default=None, help='MKLDNN (CPU)-targeted custom layers. Absolute \ path to a shared library with the kernels impl.') parser.add_argument('-u', '--utilization_monitors', default='', type=str, help='Optional. List of monitors to show initially.') args = parser.parse_args() cap = open_images_capture(args.input, args.loop) if cap.get_type() not in ('VIDEO', 'CAMERA'): raise RuntimeError("The input should be a video file or a numeric camera ID") if bool(args.m_instance_segmentation) == bool(args.m_semantic_segmentation): raise ValueError('Set up exactly one of segmentation models: ' '--m_instance_segmentation or --m_semantic_segmentation') labels_dir = Path(__file__).resolve().parents[3] / 'data/dataset_classes' mouse = MouseClick() if not args.no_show: cv2.namedWindow(WINNAME) cv2.setMouseCallback(WINNAME, mouse.get_points) log.info('OpenVINO Inference Engine') log.info('\tbuild: {}'.format(get_version())) core = Core() model_path = args.m_instance_segmentation if args.m_instance_segmentation else args.m_semantic_segmentation log.info('Reading model {}'.format(model_path)) if args.m_instance_segmentation: labels_file = str(labels_dir / 'coco_80cl_bkgr.txt') segmentation = MaskRCNN(core, args.m_instance_segmentation, labels_file, args.threshold, args.device, args.cpu_extension) elif args.m_semantic_segmentation: labels_file = str(labels_dir / 'cityscapes_19cl_bkgr.txt') segmentation = SemanticSegmentation(core, args.m_semantic_segmentation, labels_file, args.threshold, args.device, args.cpu_extension) log.info('The model {} is loaded to {}'.format(model_path, args.device)) metrics = PerformanceMetrics() video_writer = cv2.VideoWriter() black_board = False frame_number = 0 key = -1 start_time = perf_counter() frame = cap.read() if frame is None: raise RuntimeError("Can't read an image from the input") out_frame_size = (frame.shape[1], frame.shape[0] * 2) output_frame = np.full((frame.shape[0], frame.shape[1], 3), 255, dtype='uint8') presenter = monitors.Presenter(args.utilization_monitors, 20, (out_frame_size[0] // 4, out_frame_size[1] // 16)) if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), out_frame_size): raise RuntimeError("Can't open video writer") while frame is not None: mask = None detections = segmentation.get_detections([frame]) expand_mask(detections, frame.shape[1] // 27) if len(detections[0]) > 0: mask = detections[0][0][2] for i in range(1, len(detections[0])): mask = cv2.bitwise_or(mask, detections[0][i][2]) if mask is not None: mask = np.stack([mask, mask, mask], axis=-1) else: mask = np.zeros(frame.shape, dtype='uint8') clear_frame = remove_background(frame, invert_colors=not black_board) output_frame = np.where(mask, output_frame, clear_frame) merged_frame = np.vstack([frame, output_frame]) merged_frame = cv2.resize(merged_frame, out_frame_size) metrics.update(start_time, merged_frame) if video_writer.isOpened() and (args.output_limit <= 0 or frame_number <= args.output_limit-1): video_writer.write(merged_frame) presenter.drawGraphs(merged_frame) if not args.no_show: cv2.imshow(WINNAME, merged_frame) key = check_pressed_keys(key) if key == 27: # 'Esc' break if key == ord('i'): # catch pressing of key 'i' black_board = not black_board output_frame = 255 - output_frame else: presenter.handleKey(key) if mouse.crop_available: x0, x1 = min(mouse.points[0][0], mouse.points[1][0]), \ max(mouse.points[0][0], mouse.points[1][0]) y0, y1 = min(mouse.points[0][1], mouse.points[1][1]), \ max(mouse.points[0][1], mouse.points[1][1]) x1, y1 = min(x1, output_frame.shape[1] - 1), min(y1, output_frame.shape[0] - 1) board = output_frame[y0: y1, x0: x1, :] if board.shape[0] > 0 and board.shape[1] > 0: cv2.namedWindow('Board', cv2.WINDOW_KEEPRATIO) cv2.imshow('Board', board) frame_number += 1 start_time = perf_counter() frame = cap.read() metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() if args.architecture_type != 'yolov4' and args.anchors: log.warning( 'The "--anchors" option works only for "-at==yolov4". Option will be omitted' ) if args.architecture_type != 'yolov4' and args.masks: log.warning( 'The "--masks" option works only for "-at==yolov4". Option will be omitted' ) if args.architecture_type not in ['nanodet', 'nanodet-plus' ] and args.num_classes: log.warning( 'The "--num_classes" option works only for "-at==nanodet" and "-at==nanodet-plus". Option will be omitted' ) cap = open_images_capture(args.input, args.loop) if args.adapter == 'openvino': plugin_config = get_user_config(args.device, args.num_streams, args.num_threads) model_adapter = OpenvinoAdapter( create_core(), args.model, device=args.device, plugin_config=plugin_config, max_num_requests=args.num_infer_requests, model_parameters={'input_layouts': args.layout}) elif args.adapter == 'ovms': model_adapter = OVMSAdapter(args.model) configuration = { 'resize_type': args.resize_type, 'mean_values': args.mean_values, 'scale_values': args.scale_values, 'reverse_input_channels': args.reverse_input_channels, 'path_to_labels': args.labels, 'confidence_threshold': args.prob_threshold, 'input_size': args.input_size, # The CTPN specific 'num_classes': args.num_classes, # The NanoDet and NanoDetPlus specific } model = DetectionModel.create_model(args.architecture_type, model_adapter, configuration) model.log_layers_info() detector_pipeline = AsyncPipeline(model) next_frame_id = 0 next_frame_id_to_show = 0 palette = ColorPalette(len(model.labels) if model.labels else 100) metrics = PerformanceMetrics() render_metrics = PerformanceMetrics() presenter = None output_transform = None video_writer = cv2.VideoWriter() while True: if detector_pipeline.callback_exceptions: raise detector_pipeline.callback_exceptions[0] # Process all completed requests results = detector_pipeline.get_result(next_frame_id_to_show) if results: objects, frame_meta = results frame = frame_meta['frame'] start_time = frame_meta['start_time'] if len(objects) and args.raw_output_message: print_raw_results(objects, model.labels, next_frame_id_to_show) presenter.drawGraphs(frame) rendering_start_time = perf_counter() frame = draw_detections(frame, objects, palette, model.labels, output_transform) render_metrics.update(rendering_start_time) metrics.update(start_time, frame) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) next_frame_id_to_show += 1 if not args.no_show: cv2.imshow('Detection Results', frame) key = cv2.waitKey(1) ESC_KEY = 27 # Quit. if key in {ord('q'), ord('Q'), ESC_KEY}: break presenter.handleKey(key) continue if detector_pipeline.is_ready(): # Get new image/frame start_time = perf_counter() frame = cap.read() if frame is None: if next_frame_id == 0: raise ValueError("Can't read an image from the input") break if next_frame_id == 0: output_transform = OutputTransform(frame.shape[:2], args.output_resolution) if args.output_resolution: output_resolution = output_transform.new_resolution else: output_resolution = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter( args.utilization_monitors, 55, (round(output_resolution[0] / 4), round(output_resolution[1] / 8))) if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), output_resolution): raise RuntimeError("Can't open video writer") # Submit for inference detector_pipeline.submit_data(frame, next_frame_id, { 'frame': frame, 'start_time': start_time }) next_frame_id += 1 else: # Wait for empty request detector_pipeline.await_any() detector_pipeline.await_all() if detector_pipeline.callback_exceptions: raise detector_pipeline.callback_exceptions[0] # Process completed requests for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id): results = detector_pipeline.get_result(next_frame_id_to_show) objects, frame_meta = results frame = frame_meta['frame'] start_time = frame_meta['start_time'] if len(objects) and args.raw_output_message: print_raw_results(objects, model.labels, next_frame_id_to_show) presenter.drawGraphs(frame) rendering_start_time = perf_counter() frame = draw_detections(frame, objects, palette, model.labels, output_transform) render_metrics.update(rendering_start_time) metrics.update(start_time, frame) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) if not args.no_show: cv2.imshow('Detection Results', frame) key = cv2.waitKey(1) ESC_KEY = 27 # Quit. if key in {ord('q'), ord('Q'), ESC_KEY}: break presenter.handleKey(key) metrics.log_total() log_latency_per_stage(cap.reader_metrics.get_latency(), detector_pipeline.preprocess_metrics.get_latency(), detector_pipeline.inference_metrics.get_latency(), detector_pipeline.postprocess_metrics.get_latency(), render_metrics.get_latency()) for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) place_recognition = PlaceRecognition(args.model, args.device, args.gallery_folder, args.gallery_size) compute_embeddings_times = [] search_in_gallery_times = [] frames_processed = 0 presenter = monitors.Presenter(args.utilization_monitors, 0) video_writer = cv2.VideoWriter() metrics = PerformanceMetrics() while True: start_time = perf_counter() frame = cap.read() if frame is None: if frames_processed == 0: raise ValueError("Can't read an image from the input") break elapsed, probe_embedding = time_elapsed( place_recognition.compute_embedding, frame) compute_embeddings_times.append(elapsed) elapsed, (sorted_indexes, distances) = time_elapsed( place_recognition.search_in_gallery, probe_embedding) search_in_gallery_times.append(elapsed) image, key = visualize( frame, [str(place_recognition.impaths[i]) for i in sorted_indexes], distances[sorted_indexes], place_recognition.input_size, np.mean(compute_embeddings_times), np.mean(search_in_gallery_times), imshow_delay=3, presenter=presenter, no_show=args.no_show) metrics.update(start_time) if frames_processed == 0: if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), (image.shape[1], image.shape[0])): raise RuntimeError("Can't open video writer") frames_processed += 1 if video_writer.isOpened() and (args.output_limit <= 0 or frames_processed <= args.output_limit): video_writer.write(image) if key == 27: break metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) with open(args.labels, 'rt') as labels_file: class_labels = labels_file.read().splitlines() assert len(class_labels), 'The file with class labels is empty' # Plugin initialization for specified device and load extensions library if specified. log.info('OpenVINO Inference Engine') log.info('\tbuild: {}'.format(get_version())) core = Core() if args.cpu_extension and 'CPU' in args.device: core.add_extension(args.cpu_extension, 'CPU') # Read IR log.info('Reading model {}'.format(args.model)) model = core.read_model(args.model) image_input, image_info_input, ( n, c, h, w), model_type, output_names, postprocessor = check_model(model) args.no_keep_aspect_ratio = model_type == 'yolact' or args.no_keep_aspect_ratio compiled_model = core.compile_model(model, args.device) infer_request = compiled_model.create_infer_request() log.info('The model {} is loaded to {}'.format(args.model, args.device)) if args.no_track: tracker = None else: tracker = StaticIOUTracker() if args.delay: delay = args.delay else: delay = int(cap.get_type() in ('VIDEO', 'CAMERA')) frames_processed = 0 metrics = PerformanceMetrics() visualizer = Visualizer(class_labels, show_boxes=args.show_boxes, show_scores=args.show_scores) video_writer = cv2.VideoWriter() start_time = perf_counter() frame = cap.read() if frame is None: raise RuntimeError("Can't read an image from the input") out_frame_size = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter( args.utilization_monitors, 45, (round(out_frame_size[0] / 4), round(out_frame_size[1] / 8))) if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), out_frame_size): raise RuntimeError("Can't open video writer") while frame is not None: if args.no_keep_aspect_ratio: # Resize the image to a target size. scale_x = w / frame.shape[1] scale_y = h / frame.shape[0] input_image = cv2.resize(frame, (w, h)) else: # Resize the image to keep the same aspect ratio and to fit it to a window of a target size. scale_x = scale_y = min(h / frame.shape[0], w / frame.shape[1]) input_image = cv2.resize(frame, None, fx=scale_x, fy=scale_y) input_image_size = input_image.shape[:2] input_image = np.pad(input_image, ((0, h - input_image_size[0]), (0, w - input_image_size[1]), (0, 0)), mode='constant', constant_values=0) # Change data layout from HWC to CHW. input_image = input_image.transpose((2, 0, 1)) input_image = input_image.reshape((n, c, h, w)).astype(np.float32) input_image_info = np.asarray( [[input_image_size[0], input_image_size[1], 1]], dtype=np.float32) # Run the model. feed_dict = {image_input: input_image} if image_info_input: feed_dict[image_info_input] = input_image_info infer_request.infer(feed_dict) outputs = { name: infer_request.get_tensor(name).data[:] for name in output_names } # Parse detection results of the current request scores, classes, boxes, masks = postprocessor(outputs, scale_x, scale_y, *frame.shape[:2], h, w, args.prob_threshold) if len(boxes) and args.raw_output_message: log.debug( ' -------------------------- Frame # {} -------------------------- ' .format(frames_processed)) log.debug( ' Class ID | Confidence | XMIN | YMIN | XMAX | YMAX ' ) for box, cls, score in zip(boxes, classes, scores): log.debug( '{:>10} | {:>10f} | {:>8.2f} | {:>8.2f} | {:>8.2f} | {:>8.2f} ' .format(cls, score, *box)) # Get instance track IDs. masks_tracks_ids = None if tracker is not None: masks_tracks_ids = tracker(masks, classes) # Visualize masks. frame = visualizer(frame, boxes, classes, scores, presenter, masks, masks_tracks_ids) metrics.update(start_time, frame) frames_processed += 1 if video_writer.isOpened() and (args.output_limit <= 0 or frames_processed <= args.output_limit): video_writer.write(frame) if not args.no_show: # Show resulting image. cv2.imshow('Results', frame) if not args.no_show: key = cv2.waitKey(delay) esc_code = 27 if key == esc_code: break presenter.handleKey(key) start_time = perf_counter() frame = cap.read() metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) delay = int(cap.get_type() in {'VIDEO', 'CAMERA'}) if args.adapter == 'openvino': plugin_config = get_user_config(args.device, args.num_streams, args.num_threads) model_adapter = OpenvinoAdapter( create_core(), args.model, device=args.device, plugin_config=plugin_config, max_num_requests=args.num_infer_requests) elif args.adapter == 'ovms': model_adapter = OVMSAdapter(args.model) config = { 'mean_values': args.mean_values, 'scale_values': args.scale_values, 'reverse_input_channels': args.reverse_input_channels, 'topk': args.topk, 'path_to_labels': args.labels } model = Classification(model_adapter, config) model.log_layers_info() async_pipeline = AsyncPipeline(model) next_frame_id = 0 next_frame_id_to_show = 0 metrics = PerformanceMetrics() render_metrics = PerformanceMetrics() presenter = None output_transform = None video_writer = cv2.VideoWriter() ESC_KEY = 27 key = -1 while True: if async_pipeline.callback_exceptions: raise async_pipeline.callback_exceptions[0] # Process all completed requests results = async_pipeline.get_result(next_frame_id_to_show) if results: classifications, frame_meta = results frame = frame_meta['frame'] start_time = frame_meta['start_time'] if args.raw_output_message: print_raw_results(classifications, next_frame_id_to_show) presenter.drawGraphs(frame) rendering_start_time = perf_counter() frame = draw_labels(frame, classifications, output_transform) if delay or args.no_show: render_metrics.update(rendering_start_time) metrics.update(start_time, frame) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) next_frame_id_to_show += 1 if not args.no_show: cv2.imshow('Classification Results', frame) key = cv2.waitKey(delay) # Quit. if key in {ord('q'), ord('Q'), ESC_KEY}: break presenter.handleKey(key) continue if async_pipeline.is_ready(): # Get new image/frame start_time = perf_counter() frame = cap.read() if frame is None: if next_frame_id == 0: raise ValueError("Can't read an image from the input") break if next_frame_id == 0: output_transform = OutputTransform(frame.shape[:2], args.output_resolution) if args.output_resolution: output_resolution = output_transform.new_resolution else: output_resolution = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter( args.utilization_monitors, 55, (round(output_resolution[0] / 4), round(output_resolution[1] / 8))) if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), output_resolution): raise RuntimeError("Can't open video writer") # Submit for inference async_pipeline.submit_data(frame, next_frame_id, { 'frame': frame, 'start_time': start_time }) next_frame_id += 1 else: # Wait for empty request async_pipeline.await_any() async_pipeline.await_all() if key not in {ord('q'), ord('Q'), ESC_KEY}: # Process completed requests for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id): results = async_pipeline.get_result(next_frame_id_to_show) while results is None: results = async_pipeline.get_result(next_frame_id_to_show) classifications, frame_meta = results frame = frame_meta['frame'] start_time = frame_meta['start_time'] if args.raw_output_message: print_raw_results(classifications, next_frame_id_to_show) presenter.drawGraphs(frame) rendering_start_time = perf_counter() frame = draw_labels(frame, classifications, output_transform) if delay or args.no_show: render_metrics.update(rendering_start_time) metrics.update(start_time, frame) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) if not args.no_show: cv2.imshow('Classification Results', frame) key = cv2.waitKey(delay) # Quit. if key in {ord('q'), ord('Q'), ESC_KEY}: break presenter.handleKey(key) if delay or args.no_show: metrics.log_total() log_latency_per_stage(cap.reader_metrics.get_latency(), async_pipeline.preprocess_metrics.get_latency(), async_pipeline.inference_metrics.get_latency(), async_pipeline.postprocess_metrics.get_latency(), render_metrics.get_latency()) for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) if cap.get_type() not in ('VIDEO', 'CAMERA'): raise RuntimeError( "The input should be a video file or a numeric camera ID") frames = RoiDetectorOnVideo(cap) img_retrieval = ImageRetrieval(args.model, args.device, args.gallery, INPUT_SIZE, args.cpu_extension) compute_embeddings_times = [] search_in_gallery_times = [] positions = [] frames_processed = 0 presenter = monitors.Presenter(args.utilization_monitors, 0) video_writer = cv2.VideoWriter() metrics = PerformanceMetrics() for image, view_frame in frames: start_time = perf_counter() position = None sorted_indexes = [] if image is not None: image = central_crop(image, divide_by=5, shift=1) elapsed, probe_embedding = time_elapsed( img_retrieval.compute_embedding, image) compute_embeddings_times.append(elapsed) elapsed, (sorted_indexes, distances) = time_elapsed( img_retrieval.search_in_gallery, probe_embedding) search_in_gallery_times.append(elapsed) sorted_classes = [ img_retrieval.gallery_classes[i] for i in sorted_indexes ] if args.ground_truth is not None: position = sorted_classes.index( img_retrieval.text_label_to_class_id[args.ground_truth]) positions.append(position) log.info("ROI detected, found: %d, position of target: %d", sorted_classes[0], position) else: log.info("ROI detected, found: %s", sorted_classes[0]) image, key = visualize( view_frame, position, [img_retrieval.impaths[i] for i in sorted_indexes], distances[sorted_indexes] if position is not None else None, img_retrieval.input_size, np.mean(compute_embeddings_times), np.mean(search_in_gallery_times), imshow_delay=3, presenter=presenter, no_show=args.no_show) metrics.update(start_time) if frames_processed == 0: if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), (image.shape[1], image.shape[0])): raise RuntimeError("Can't open video writer") frames_processed += 1 if video_writer.isOpened() and (args.output_limit <= 0 or frames_processed <= args.output_limit): video_writer.write(image) if key == 27: break metrics.log_total() for rep in presenter.reportMeans(): log.info(rep) if positions: compute_metrics(positions)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) target_bgr = open_images_capture(args.target_bgr, loop=True) if args.target_bgr else None if args.adapter == 'openvino': plugin_config = get_user_config(args.device, args.num_streams, args.num_threads) model_adapter = OpenvinoAdapter(create_core(), args.model, device=args.device, plugin_config=plugin_config, max_num_requests=args.num_infer_requests) elif args.adapter == 'ovms': model_adapter = OVMSAdapter(args.model) labels = ['__background__', 'person'] if args.labels is None else load_labels(args.labels) assert len(labels), 'The file with class labels is empty' configuration = { 'confidence_threshold': args.prob_threshold, 'resize_type': args.resize_type } model, need_bgr_input = get_model(model_adapter, configuration, args) input_bgr = open_images_capture(args.background, False).read() if need_bgr_input else None person_id = -1 for i, label in enumerate(labels): if label == 'person': person_id = i break assert person_id >= 0, 'Person class did not find in labels list.' model.log_layers_info() pipeline = AsyncPipeline(model) next_frame_id = 0 next_frame_id_to_show = 0 metrics = PerformanceMetrics() render_metrics = PerformanceMetrics() presenter = None output_transform = None video_writer = cv2.VideoWriter() while True: if pipeline.is_ready(): # Get new image/frame start_time = perf_counter() frame = cap.read() bgr = target_bgr.read() if target_bgr is not None else None if frame is None: if next_frame_id == 0: raise ValueError("Can't read an image from the input") break if next_frame_id == 0: output_transform = OutputTransform(frame.shape[:2], args.output_resolution) if args.output_resolution: output_resolution = output_transform.new_resolution else: output_resolution = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter(args.utilization_monitors, 55, (round(output_resolution[0] / 4), round(output_resolution[1] / 8))) if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), tuple(output_resolution)): raise RuntimeError("Can't open video writer") # Submit for inference data = {'src': frame, 'bgr': input_bgr} if input_bgr is not None else frame pipeline.submit_data(data, next_frame_id, {'frame': frame, 'start_time': start_time}) next_frame_id += 1 else: # Wait for empty request pipeline.await_any() if pipeline.callback_exceptions: raise pipeline.callback_exceptions[0] # Process all completed requests results = pipeline.get_result(next_frame_id_to_show) if results: objects, frame_meta = results if args.raw_output_message: print_raw_results(objects, next_frame_id_to_show) frame = frame_meta['frame'] start_time = frame_meta['start_time'] rendering_start_time = perf_counter() frame = render_results(frame, objects, output_resolution, bgr, person_id, args.blur_bgr, args.show_with_original_frame) render_metrics.update(rendering_start_time) presenter.drawGraphs(frame) metrics.update(start_time, frame) if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit-1): video_writer.write(frame) next_frame_id_to_show += 1 if not args.no_show: cv2.imshow('Background subtraction results', frame) key = cv2.waitKey(1) if key == 27 or key == 'q' or key == 'Q': break presenter.handleKey(key) pipeline.await_all() # Process completed requests for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id): results = pipeline.get_result(next_frame_id_to_show) while results is None: results = pipeline.get_result(next_frame_id_to_show) objects, frame_meta = results if args.raw_output_message: print_raw_results(objects, next_frame_id_to_show, model.labels) frame = frame_meta['frame'] start_time = frame_meta['start_time'] rendering_start_time = perf_counter() frame = render_results(frame, objects, output_resolution, bgr, person_id, args.blur_bgr, args.show_with_original_frame) render_metrics.update(rendering_start_time) presenter.drawGraphs(frame) metrics.update(start_time, frame) if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit-1): video_writer.write(frame) if not args.no_show: cv2.imshow('Background subtraction results', frame) cv2.waitKey(1) metrics.log_total() log_latency_per_stage(cap.reader_metrics.get_latency(), pipeline.preprocess_metrics.get_latency(), pipeline.inference_metrics.get_latency(), pipeline.postprocess_metrics.get_latency(), render_metrics.get_latency()) for rep in presenter.reportMeans(): log.info(rep)
R = np.array(extrinsics['R'], dtype=np.float32) t = np.array(extrinsics['t'], dtype=np.float32) is_video = cap.get_type() in ('VIDEO', 'CAMERA') base_height = args.height_size fx = args.fx frames_processed = 0 delay = 1 esc_code = 27 p_code = 112 space_code = 32 mean_time = 0 presenter = monitors.Presenter(args.utilization_monitors, 0) metrics = PerformanceMetrics() video_writer = cv2.VideoWriter() start_time = perf_counter() frame = cap.read() if frame is None: raise RuntimeError("Can't read an image from the input") if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), (frame.shape[1], frame.shape[0])): raise RuntimeError("Can't open video writer") while frame is not None: current_time = cv2.getTickCount() input_scale = base_height / frame.shape[0]
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) if args.adapter == 'openvino': plugin_config = get_user_config(args.device, args.num_streams, args.num_threads) model_adapter = OpenvinoAdapter( create_core(), args.model, device=args.device, plugin_config=plugin_config, max_num_requests=args.num_infer_requests, model_parameters={'input_layouts': args.layout}) elif args.adapter == 'ovms': model_adapter = OVMSAdapter(args.model) model = SegmentationModel.create_model(args.architecture_type, model_adapter, {'path_to_labels': args.labels}) if args.architecture_type == 'segmentation': visualizer = SegmentationVisualizer(args.colors) if args.architecture_type == 'salient_object_detection': visualizer = SaliencyMapVisualizer() model.log_layers_info() pipeline = AsyncPipeline(model) next_frame_id = 0 next_frame_id_to_show = 0 metrics = PerformanceMetrics() render_metrics = PerformanceMetrics() presenter = None output_transform = None video_writer = cv2.VideoWriter() only_masks = args.only_masks while True: if pipeline.is_ready(): # Get new image/frame start_time = perf_counter() frame = cap.read() if frame is None: if next_frame_id == 0: raise ValueError("Can't read an image from the input") break if next_frame_id == 0: output_transform = OutputTransform(frame.shape[:2], args.output_resolution) if args.output_resolution: output_resolution = output_transform.new_resolution else: output_resolution = (frame.shape[1], frame.shape[0]) presenter = monitors.Presenter( args.utilization_monitors, 55, (round(output_resolution[0] / 4), round(output_resolution[1] / 8))) if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), output_resolution): raise RuntimeError("Can't open video writer") # Submit for inference pipeline.submit_data(frame, next_frame_id, { 'frame': frame, 'start_time': start_time }) next_frame_id += 1 else: # Wait for empty request pipeline.await_any() if pipeline.callback_exceptions: raise pipeline.callback_exceptions[0] # Process all completed requests results = pipeline.get_result(next_frame_id_to_show) if results: objects, frame_meta = results if args.raw_output_message: print_raw_results(objects, next_frame_id_to_show, model.labels) frame = frame_meta['frame'] start_time = frame_meta['start_time'] rendering_start_time = perf_counter() frame = render_segmentation(frame, objects, visualizer, output_transform, only_masks) render_metrics.update(rendering_start_time) presenter.drawGraphs(frame) metrics.update(start_time, frame) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) next_frame_id_to_show += 1 if not args.no_show: cv2.imshow('Segmentation Results', frame) key = cv2.waitKey(1) if key == 27 or key == 'q' or key == 'Q': break if key == 9: only_masks = not only_masks presenter.handleKey(key) pipeline.await_all() if pipeline.callback_exceptions: raise pipeline.callback_exceptions[0] # Process completed requests for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id): results = pipeline.get_result(next_frame_id_to_show) objects, frame_meta = results if args.raw_output_message: print_raw_results(objects, next_frame_id_to_show, model.labels) frame = frame_meta['frame'] start_time = frame_meta['start_time'] rendering_start_time = perf_counter() frame = render_segmentation(frame, objects, visualizer, output_transform, only_masks) render_metrics.update(rendering_start_time) presenter.drawGraphs(frame) metrics.update(start_time, frame) if video_writer.isOpened() and ( args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1): video_writer.write(frame) if not args.no_show: cv2.imshow('Segmentation Results', frame) key = cv2.waitKey(1) metrics.log_total() log_latency_per_stage(cap.reader_metrics.get_latency(), pipeline.preprocess_metrics.get_latency(), pipeline.inference_metrics.get_latency(), pipeline.postprocess_metrics.get_latency(), render_metrics.get_latency()) for rep in presenter.reportMeans(): log.info(rep)
def main(args): cap = open_images_capture(args.input, args.loop) log.info('OpenVINO Inference Engine') log.info('\tbuild: {}'.format(get_version())) core = Core() log.info('Reading model {}'.format(args.model)) model = core.read_model(args.model) input_tensor_name = 'data_l' input_shape = model.input(input_tensor_name).shape assert input_shape[1] == 1, "Expected model input shape with 1 channel" inputs = {} for input in model.inputs: inputs[input.get_any_name()] = np.zeros(input.shape) assert len(model.outputs) == 1, "Expected number of outputs is equal 1" compiled_model = core.compile_model(model, device_name=args.device) output_tensor = compiled_model.outputs[0] infer_request = compiled_model.create_infer_request() log.info('The model {} is loaded to {}'.format(args.model, args.device)) _, _, h_in, w_in = input_shape frames_processed = 0 imshow_size = (640, 480) graph_size = (imshow_size[0] // 2, imshow_size[1] // 4) presenter = monitors.Presenter(args.utilization_monitors, imshow_size[1] * 2 - graph_size[1], graph_size) metrics = PerformanceMetrics() video_writer = cv.VideoWriter() if args.output and not video_writer.open( args.output, cv.VideoWriter_fourcc(*'MJPG'), cap.fps(), (imshow_size[0] * 2, imshow_size[1] * 2)): raise RuntimeError("Can't open video writer") start_time = perf_counter() original_frame = cap.read() if original_frame is None: raise RuntimeError("Can't read an image from the input") while original_frame is not None: (h_orig, w_orig) = original_frame.shape[:2] if original_frame.shape[2] > 1: frame = cv.cvtColor(cv.cvtColor(original_frame, cv.COLOR_BGR2GRAY), cv.COLOR_GRAY2RGB) else: frame = cv.cvtColor(original_frame, cv.COLOR_GRAY2RGB) img_rgb = frame.astype(np.float32) / 255 img_lab = cv.cvtColor(img_rgb, cv.COLOR_RGB2Lab) img_l_rs = cv.resize(img_lab.copy(), (w_in, h_in))[:, :, 0] inputs[input_tensor_name] = np.expand_dims(img_l_rs, axis=[0, 1]) res = infer_request.infer(inputs)[output_tensor] update_res = np.squeeze(res) out = update_res.transpose((1, 2, 0)) out = cv.resize(out, (w_orig, h_orig)) img_lab_out = np.concatenate((img_lab[:, :, 0][:, :, np.newaxis], out), axis=2) img_bgr_out = np.clip(cv.cvtColor(img_lab_out, cv.COLOR_Lab2BGR), 0, 1) original_image = cv.resize(original_frame, imshow_size) grayscale_image = cv.resize(frame, imshow_size) colorize_image = (cv.resize(img_bgr_out, imshow_size) * 255).astype( np.uint8) lab_image = cv.resize(img_lab_out, imshow_size).astype(np.uint8) original_image = cv.putText(original_image, 'Original', (25, 50), cv.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2, cv.LINE_AA) grayscale_image = cv.putText(grayscale_image, 'Grayscale', (25, 50), cv.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2, cv.LINE_AA) colorize_image = cv.putText(colorize_image, 'Colorize', (25, 50), cv.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2, cv.LINE_AA) lab_image = cv.putText(lab_image, 'LAB interpretation', (25, 50), cv.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2, cv.LINE_AA) ir_image = [ cv.hconcat([original_image, grayscale_image]), cv.hconcat([lab_image, colorize_image]) ] final_image = cv.vconcat(ir_image) metrics.update(start_time, final_image) frames_processed += 1 if video_writer.isOpened() and (args.output_limit <= 0 or frames_processed <= args.output_limit): video_writer.write(final_image) presenter.drawGraphs(final_image) if not args.no_show: cv.imshow('Colorization Demo', final_image) key = cv.waitKey(1) if key in {ord("q"), ord("Q"), 27}: break presenter.handleKey(key) start_time = perf_counter() original_frame = cap.read() metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)
def main(): args = build_argparser().parse_args() cap = open_images_capture(args.input, args.loop) log.info('OpenVINO Inference Engine') log.info('\tbuild: {}'.format(get_version())) core = Core() # Read IR log.info('Reading Mask-RCNN model {}'.format(args.mask_rcnn_model)) mask_rcnn_model = core.read_model(args.mask_rcnn_model) input_tensor_name = 'image' try: n, c, h, w = mask_rcnn_model.input(input_tensor_name).shape if n != 1: raise RuntimeError( 'Only batch 1 is supported by the demo application') except RuntimeError: raise RuntimeError( 'Demo supports only topologies with the following input tensor name: {}' .format(input_tensor_name)) required_output_names = {'boxes', 'labels', 'masks', 'text_features'} for output_tensor_name in required_output_names: try: mask_rcnn_model.output(output_tensor_name) except RuntimeError: raise RuntimeError( 'Demo supports only topologies with the following output tensor names: {}' .format(', '.join(required_output_names))) log.info('Reading Text Recognition Encoder model {}'.format( args.text_enc_model)) text_enc_model = core.read_model(args.text_enc_model) log.info('Reading Text Recognition Decoder model {}'.format( args.text_dec_model)) text_dec_model = core.read_model(args.text_dec_model) mask_rcnn_compiled_model = core.compile_model(mask_rcnn_model, device_name=args.device) mask_rcnn_infer_request = mask_rcnn_compiled_model.create_infer_request() log.info('The Mask-RCNN model {} is loaded to {}'.format( args.mask_rcnn_model, args.device)) text_enc_compiled_model = core.compile_model(text_enc_model, args.device) text_enc_output_tensor = text_enc_compiled_model.outputs[0] text_enc_infer_request = text_enc_compiled_model.create_infer_request() log.info('The Text Recognition Encoder model {} is loaded to {}'.format( args.text_enc_model, args.device)) text_dec_compiled_model = core.compile_model(text_dec_model, args.device) text_dec_infer_request = text_dec_compiled_model.create_infer_request() log.info('The Text Recognition Decoder model {} is loaded to {}'.format( args.text_dec_model, args.device)) hidden_shape = text_dec_model.input(args.trd_input_prev_hidden).shape text_dec_output_names = { args.trd_output_symbols_distr, args.trd_output_cur_hidden } if args.no_track: tracker = None else: tracker = StaticIOUTracker() if args.delay: delay = args.delay else: delay = int(cap.get_type() in ('VIDEO', 'CAMERA')) visualizer = InstanceSegmentationVisualizer(show_boxes=args.show_boxes, show_scores=args.show_scores) frames_processed = 0 metrics = PerformanceMetrics() video_writer = cv2.VideoWriter() start_time = perf_counter() frame = cap.read() if frame is None: raise RuntimeError("Can't read an image from the input") presenter = monitors.Presenter(args.utilization_monitors, 45, (frame.shape[1] // 4, frame.shape[0] // 8)) if args.output and not video_writer.open( args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(), (frame.shape[1], frame.shape[0])): raise RuntimeError("Can't open video writer") while frame is not None: if not args.keep_aspect_ratio: # Resize the image to a target size. scale_x = w / frame.shape[1] scale_y = h / frame.shape[0] input_image = cv2.resize(frame, (w, h)) else: # Resize the image to keep the same aspect ratio and to fit it to a window of a target size. scale_x = scale_y = min(h / frame.shape[0], w / frame.shape[1]) input_image = cv2.resize(frame, None, fx=scale_x, fy=scale_y) input_image_size = input_image.shape[:2] input_image = np.pad(input_image, ((0, h - input_image_size[0]), (0, w - input_image_size[1]), (0, 0)), mode='constant', constant_values=0) # Change data layout from HWC to CHW. input_image = input_image.transpose((2, 0, 1)) input_image = input_image.reshape((n, c, h, w)).astype(np.float32) # Run the MaskRCNN model. mask_rcnn_infer_request.infer({input_tensor_name: input_image}) outputs = { name: mask_rcnn_infer_request.get_tensor(name).data[:] for name in required_output_names } # Parse detection results of the current request boxes = outputs['boxes'][:, :4] scores = outputs['boxes'][:, 4] classes = outputs['labels'].astype(np.uint32) raw_masks = outputs['masks'] text_features = outputs['text_features'] # Filter out detections with low confidence. detections_filter = scores > args.prob_threshold scores = scores[detections_filter] classes = classes[detections_filter] boxes = boxes[detections_filter] raw_masks = raw_masks[detections_filter] text_features = text_features[detections_filter] boxes[:, 0::2] /= scale_x boxes[:, 1::2] /= scale_y masks = [] for box, cls, raw_mask in zip(boxes, classes, raw_masks): mask = segm_postprocess(box, raw_mask, frame.shape[0], frame.shape[1]) masks.append(mask) texts = [] for feature in text_features: input_data = {'input': np.expand_dims(feature, axis=0)} feature = text_enc_infer_request.infer( input_data)[text_enc_output_tensor] feature = np.reshape(feature, (feature.shape[0], feature.shape[1], -1)) feature = np.transpose(feature, (0, 2, 1)) hidden = np.zeros(hidden_shape) prev_symbol_index = np.ones((1, )) * SOS_INDEX text = '' text_confidence = 1.0 for i in range(MAX_SEQ_LEN): text_dec_infer_request.infer({ args.trd_input_prev_symbol: np.reshape(prev_symbol_index, (1, )), args.trd_input_prev_hidden: hidden, args.trd_input_encoder_outputs: feature }) decoder_output = { name: text_dec_infer_request.get_tensor(name).data[:] for name in text_dec_output_names } symbols_distr = decoder_output[args.trd_output_symbols_distr] symbols_distr_softmaxed = softmax(symbols_distr, axis=1)[0] prev_symbol_index = int(np.argmax(symbols_distr, axis=1)) text_confidence *= symbols_distr_softmaxed[prev_symbol_index] if prev_symbol_index == EOS_INDEX: break text += args.alphabet[prev_symbol_index] hidden = decoder_output[args.trd_output_cur_hidden] texts.append(text if text_confidence >= args.tr_threshold else '') if len(boxes) and args.raw_output_message: log.debug( ' -------------------------- Frame # {} -------------------------- ' .format(frames_processed)) log.debug( ' Class ID | Confidence | XMIN | YMIN | XMAX | YMAX ' ) for box, cls, score, mask in zip(boxes, classes, scores, masks): log.debug( '{:>10} | {:>10f} | {:>8.2f} | {:>8.2f} | {:>8.2f} | {:>8.2f} ' .format(cls, score, *box)) # Get instance track IDs. masks_tracks_ids = None if tracker is not None: masks_tracks_ids = tracker(masks, classes) presenter.drawGraphs(frame) # Visualize masks. frame = visualizer(frame, boxes, classes, scores, masks, masks_tracks_ids, texts) metrics.update(start_time, frame) frames_processed += 1 if video_writer.isOpened() and (args.output_limit <= 0 or frames_processed <= args.output_limit): video_writer.write(frame) if not args.no_show: # Show resulting image. cv2.imshow('Results', frame) if not args.no_show: key = cv2.waitKey(delay) esc_code = 27 if key == esc_code: break presenter.handleKey(key) start_time = perf_counter() frame = cap.read() metrics.log_total() for rep in presenter.reportMeans(): log.info(rep)