class Detector(object): def __init__(self, opt): if opt.gpus[0] >= 0: opt.device = torch.device('cuda') else: opt.device = torch.device('cpu') print('Creating model...') self.model = create_model(opt.arch, opt.heads, opt.head_conv, opt=opt) self.model = load_model(self.model, opt.load_model, opt) self.model = self.model.to(opt.device) self.model.eval() self.opt = opt self.trained_dataset = get_dataset(opt.dataset) self.mean = np.array(self.trained_dataset.mean, dtype=np.float32).reshape(1, 1, 3) self.std = np.array(self.trained_dataset.std, dtype=np.float32).reshape(1, 1, 3) self.pause = not opt.no_pause self.rest_focal_length = self.trained_dataset.rest_focal_length \ if self.opt.test_focal_length < 0 else self.opt.test_focal_length self.flip_idx = self.trained_dataset.flip_idx self.cnt = 0 self.pre_images = None self.pre_image_ori = None self.tracker = Tracker(opt) self.debugger = Debugger(opt=opt, dataset=self.trained_dataset) def run(self, image_or_path_or_tensor, meta={}): load_time, pre_time, net_time, dec_time, post_time = 0, 0, 0, 0, 0 merge_time, track_time, tot_time, display_time = 0, 0, 0, 0 self.debugger.clear() start_time = time.time() # read image pre_processed = False if isinstance(image_or_path_or_tensor, np.ndarray): image = image_or_path_or_tensor elif type(image_or_path_or_tensor) == type(''): image = cv2.imread(image_or_path_or_tensor) else: image = image_or_path_or_tensor['image'][0].numpy() pre_processed_images = image_or_path_or_tensor pre_processed = True loaded_time = time.time() load_time += (loaded_time - start_time) detections = [] # for multi-scale testing for scale in self.opt.test_scales: scale_start_time = time.time() if not pre_processed: # not prefetch testing or demo images, meta = self.pre_process(image, scale, meta) else: # prefetch testing images = pre_processed_images['images'][scale][0] meta = pre_processed_images['meta'][scale] meta = {k: v.numpy()[0] for k, v in meta.items()} if 'pre_dets' in pre_processed_images['meta']: meta['pre_dets'] = pre_processed_images['meta']['pre_dets'] if 'cur_dets' in pre_processed_images['meta']: meta['cur_dets'] = pre_processed_images['meta']['cur_dets'] images = images.to(self.opt.device, non_blocking=self.opt.non_block_test) # initializing tracker pre_hms, pre_inds = None, None if self.opt.tracking: # initialize the first frame if self.pre_images is None: print('Initialize tracking!') self.pre_images = images self.tracker.init_track(meta['pre_dets'] if 'pre_dets' in meta else []) if self.opt.pre_hm: # render input heatmap from tracker status # pre_inds is not used in the current version. # We used pre_inds for learning an offset from previous image to # the current image. pre_hms, pre_inds = self._get_additional_inputs( self.tracker.tracks, meta, with_hm=not self.opt.zero_pre_hm) pre_process_time = time.time() pre_time += pre_process_time - scale_start_time # run the network # output: the output feature maps, only used for visualizing # dets: output tensors after extracting peaks output, dets, forward_time = self.process(images, self.pre_images, pre_hms, pre_inds, return_time=True) net_time += forward_time - pre_process_time decode_time = time.time() dec_time += decode_time - forward_time # convert the cropped and 4x downsampled output coordinate system # back to the input image coordinate system result = self.post_process(dets, meta, scale) post_process_time = time.time() post_time += post_process_time - decode_time detections.append(result) if self.opt.debug >= 2: self.debug(self.debugger, images, result, output, scale, pre_images=self.pre_images if not self.opt.no_pre_img else None, pre_hms=pre_hms) # merge multi-scale testing results results = self.merge_outputs(detections) if self.opt.gpus[0] >= 0: torch.cuda.synchronize() end_time = time.time() merge_time += end_time - post_process_time if self.opt.tracking: # public detection mode in MOT challenge public_det = meta['cur_dets'] if self.opt.public_det else None # add tracking id to results results = self.tracker.step(results, public_det) self.pre_images = images tracking_time = time.time() track_time += tracking_time - end_time tot_time += tracking_time - start_time if self.opt.debug >= 1: self.show_results(self.debugger, image, results) self.cnt += 1 show_results_time = time.time() display_time += show_results_time - end_time # return results and run time ret = { 'results': results, 'tot': tot_time, 'load': load_time, 'pre': pre_time, 'net': net_time, 'dec': dec_time, 'post': post_time, 'merge': merge_time, 'track': track_time, 'display': display_time } if self.opt.save_video: try: # return debug image for saving video ret.update({'generic': self.debugger.imgs['generic']}) except: pass return ret def _transform_scale(self, image, scale=1): ''' Prepare input image in different testing modes. Currently support: fix short size/ center crop to a fixed size/ keep original resolution but pad to a multiplication of 32 ''' height, width = image.shape[0:2] new_height = int(height * scale) new_width = int(width * scale) if self.opt.fix_short > 0: if height < width: inp_height = self.opt.fix_short inp_width = (int(width / height * self.opt.fix_short) + 63) // 64 * 64 else: inp_height = (int(height / width * self.opt.fix_short) + 63) // 64 * 64 inp_width = self.opt.fix_short c = np.array([width / 2, height / 2], dtype=np.float32) s = np.array([width, height], dtype=np.float32) elif self.opt.fix_res: inp_height, inp_width = self.opt.input_h, self.opt.input_w c = np.array([new_width / 2., new_height / 2.], dtype=np.float32) s = max(height, width) * 1.0 # s = np.array([inp_width, inp_height], dtype=np.float32) else: inp_height = (new_height | self.opt.pad) + 1 inp_width = (new_width | self.opt.pad) + 1 c = np.array([new_width // 2, new_height // 2], dtype=np.float32) s = np.array([inp_width, inp_height], dtype=np.float32) resized_image = cv2.resize(image, (new_width, new_height)) return resized_image, c, s, inp_width, inp_height, height, width def pre_process(self, image, scale, input_meta={}): ''' Crop, resize, and normalize image. Gather meta data for post processing and tracking. ''' resized_image, c, s, inp_width, inp_height, height, width = \ self._transform_scale(image) trans_input = get_affine_transform(c, s, 0, [inp_width, inp_height]) out_height = inp_height // self.opt.down_ratio out_width = inp_width // self.opt.down_ratio trans_output = get_affine_transform(c, s, 0, [out_width, out_height]) inp_image = cv2.warpAffine(resized_image, trans_input, (inp_width, inp_height), flags=cv2.INTER_LINEAR) inp_image = ((inp_image / 255. - self.mean) / self.std).astype( np.float32) images = inp_image.transpose(2, 0, 1).reshape(1, 3, inp_height, inp_width) if self.opt.flip_test: images = np.concatenate((images, images[:, :, :, ::-1]), axis=0) images = torch.from_numpy(images) meta = {'calib': np.array(input_meta['calib'], dtype=np.float32) \ if 'calib' in input_meta else \ self._get_default_calib(width, height)} meta.update({ 'c': c, 's': s, 'height': height, 'width': width, 'out_height': out_height, 'out_width': out_width, 'inp_height': inp_height, 'inp_width': inp_width, 'trans_input': trans_input, 'trans_output': trans_output }) if 'pre_dets' in input_meta: meta['pre_dets'] = input_meta['pre_dets'] if 'cur_dets' in input_meta: meta['cur_dets'] = input_meta['cur_dets'] return images, meta def _trans_bbox(self, bbox, trans, width, height): ''' Transform bounding boxes according to image crop. ''' bbox = np.array(copy.deepcopy(bbox), dtype=np.float32) bbox[:2] = affine_transform(bbox[:2], trans) bbox[2:] = affine_transform(bbox[2:], trans) bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, width - 1) bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, height - 1) return bbox def _get_additional_inputs(self, dets, meta, with_hm=True): ''' Render input heatmap from previous trackings. ''' trans_input, trans_output = meta['trans_input'], meta['trans_output'] inp_width, inp_height = meta['inp_width'], meta['inp_height'] out_width, out_height = meta['out_width'], meta['out_height'] input_hm = np.zeros((1, inp_height, inp_width), dtype=np.float32) output_inds = [] for det in dets: if det['score'] < self.opt.pre_thresh or det['active'] == 0: continue bbox = self._trans_bbox(det['bbox'], trans_input, inp_width, inp_height) bbox_out = self._trans_bbox(det['bbox'], trans_output, out_width, out_height) h, w = bbox[3] - bbox[1], bbox[2] - bbox[0] if (h > 0 and w > 0): radius = gaussian_radius((math.ceil(h), math.ceil(w))) radius = max(0, int(radius)) ct = np.array([(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2], dtype=np.float32) ct_int = ct.astype(np.int32) if with_hm: draw_umich_gaussian(input_hm[0], ct_int, radius) ct_out = np.array([(bbox_out[0] + bbox_out[2]) / 2, (bbox_out[1] + bbox_out[3]) / 2], dtype=np.int32) output_inds.append(ct_out[1] * out_width + ct_out[0]) if with_hm: input_hm = input_hm[np.newaxis] if self.opt.flip_test: input_hm = np.concatenate((input_hm, input_hm[:, :, :, ::-1]), axis=0) input_hm = torch.from_numpy(input_hm).to(self.opt.device) output_inds = np.array(output_inds, np.int64).reshape(1, -1) output_inds = torch.from_numpy(output_inds).to(self.opt.device) return input_hm, output_inds def _get_default_calib(self, width, height): calib = np.array([[self.rest_focal_length, 0, width / 2, 0], [0, self.rest_focal_length, height / 2, 0], [0, 0, 1, 0]]) return calib def _sigmoid_output(self, output): if 'hm' in output: output['hm'] = output['hm'].sigmoid_() if 'hm_hp' in output: output['hm_hp'] = output['hm_hp'].sigmoid_() if 'dep' in output: output['dep'] = 1. / (output['dep'].sigmoid() + 1e-6) - 1. output['dep'] *= self.opt.depth_scale return output def _flip_output(self, output): average_flips = ['hm', 'wh', 'dep', 'dim'] ##TODO consider tracking_wh neg_average_flips = ['amodel_offset'] single_flips = [ 'ltrb', 'nuscenes_att', 'velocity', 'ltrb_amodal', 'reg', 'hp_offset', 'rot', 'tracking', 'pre_hm' ] ## TODO consider iou for head in output: if head in average_flips: output[head] = (output[head][0:1] + flip_tensor(output[head][1:2])) / 2 if head in neg_average_flips: flipped_tensor = flip_tensor(output[head][1:2]) flipped_tensor[:, 0::2] *= -1 output[head] = (output[head][0:1] + flipped_tensor) / 2 if head in single_flips: output[head] = output[head][0:1] if head == 'hps': output['hps'] = (output['hps'][0:1] + flip_lr_off( output['hps'][1:2], self.flip_idx)) / 2 if head == 'hm_hp': output['hm_hp'] = (output['hm_hp'][0:1] + \ flip_lr(output['hm_hp'][1:2], self.flip_idx)) / 2 return output def process(self, images, pre_images=None, pre_hms=None, pre_inds=None, return_time=False): with torch.no_grad(): if self.opt.gpus[0] >= 0: torch.cuda.synchronize() output = self.model(images, pre_images, pre_hms)[-1] output = self._sigmoid_output(output) output.update({'pre_inds': pre_inds}) if self.opt.flip_test: output = self._flip_output(output) if self.opt.gpus[0] >= 0: torch.cuda.synchronize() forward_time = time.time() dets = generic_decode(output, K=self.opt.K, opt=self.opt) if self.opt.gpus[0] >= 0: torch.cuda.synchronize() for k in dets: dets[k] = dets[k].detach().cpu().numpy() if return_time: return output, dets, forward_time else: return output, dets def post_process(self, dets, meta, scale=1): dets = generic_post_process(self.opt, dets, [meta['c']], [meta['s']], meta['out_height'], meta['out_width'], self.opt.num_classes, [meta['calib']], meta['height'], meta['width']) self.this_calib = meta['calib'] if scale != 1: for i in range(len(dets[0])): for k in ['bbox', 'hps']: if k in dets[0][i]: dets[0][i][k] = (np.array(dets[0][i][k], np.float32) / scale).tolist() return dets[0] def merge_outputs(self, detections): assert len(self.opt.test_scales) == 1, 'multi_scale not supported!' results = [] for i in range(len(detections[0])): if detections[0][i]['score'] > self.opt.out_thresh: results.append(detections[0][i]) return results def debug(self, debugger, images, dets, output, scale=1, pre_images=None, pre_hms=None): img = images[0].detach().cpu().numpy().transpose(1, 2, 0) img = np.clip(((img * self.std + self.mean) * 255.), 0, 255).astype(np.uint8) pred = debugger.gen_colormap(output['hm'][0].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hm') if 'hm_hp' in output: pred = debugger.gen_colormap_hp( output['hm_hp'][0].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hmhp') if pre_images is not None: pre_img = pre_images[0].detach().cpu().numpy().transpose(1, 2, 0) pre_img = np.clip(((pre_img * self.std + self.mean) * 255.), 0, 255).astype(np.uint8) debugger.add_img(pre_img, 'pre_img') if pre_hms is not None: pre_hm = debugger.gen_colormap( pre_hms[0].detach().cpu().numpy()) debugger.add_blend_img(pre_img, pre_hm, 'pre_hm') def show_results(self, debugger, image, results): debugger.add_img(image, img_id='generic') # if self.opt.tracking: # debugger.add_img(self.pre_image_ori if self.pre_image_ori is not None else image, # img_id='previous') # self.pre_image_ori = image for j in range(len(results)): if results[j]['score'] > self.opt.vis_thresh: if 'active' in results[j] and results[j]['active'] == 0: continue item = results[j] if ('bbox' in item): sc = item['score'] if self.opt.demo == '' or \ not ('tracking_id' in item) else item['tracking_id'] sc = item[ 'tracking_id'] if self.opt.show_track_color else sc debugger.add_coco_bbox(item['bbox'], item['class'] - 1, sc, img_id='generic') if 'tracking' in item: debugger.add_arrow(item['ct'], item['tracking'], img_id='generic') tracking_id = item[ 'tracking_id'] if 'tracking_id' in item else -1 if 'tracking_id' in item and self.opt.demo == '' and \ not self.opt.show_track_color: debugger.add_tracking_id(item['ct'], item['tracking_id'], img_id='generic') if (item['class'] in [1, 2]) and 'hps' in item: debugger.add_coco_hp(item['hps'], tracking_id=tracking_id, img_id='generic') if len(results) > 0 and \ 'dep' in results[0] and 'alpha' in results[0] and 'dim' in results[0]: debugger.add_3d_detection( image if not self.opt.qualitative else cv2.resize( debugger.imgs['pred_hm'], (image.shape[1], image.shape[0])), False, results, self.this_calib, vis_thresh=self.opt.vis_thresh, img_id='ddd_pred') debugger.add_bird_view(results, vis_thresh=self.opt.vis_thresh, img_id='bird_pred', cnt=self.cnt) if self.opt.show_track_color and self.opt.debug == 4: del debugger.imgs['generic'], debugger.imgs['bird_pred'] if 'ddd_pred' in debugger.imgs: debugger.imgs['generic'] = debugger.imgs['ddd_pred'] if self.opt.debug == 4: debugger.save_all_imgs(self.opt.debug_dir, prefix='{}'.format(self.cnt)) else: debugger.show_all_imgs(pause=self.pause) def reset_tracking(self): self.tracker.reset() self.pre_images = None self.pre_image_ori = None
class Detector(object): def __init__(self, opt): if opt.gpus[0] >= 0: opt.device = torch.device("cuda") else: opt.device = torch.device("cpu") print("Creating model...") self.model = create_model(opt.arch, opt.heads, opt.head_conv, opt=opt) self.model = load_model(self.model, opt.load_model, opt) self.model = self.model.to(opt.device) self.model.eval() self.opt = opt self.trained_dataset = get_dataset(opt.dataset) self.mean = np.array(self.trained_dataset.mean, dtype=np.float32).reshape(1, 1, 3) self.std = np.array(self.trained_dataset.std, dtype=np.float32).reshape(1, 1, 3) # self.pause = not opt.no_pause self.rest_focal_length = (self.trained_dataset.rest_focal_length if self.opt.test_focal_length < 0 else self.opt.test_focal_length) self.flip_idx = self.trained_dataset.flip_idx self.cnt = 0 self.pre_images = None self.pre_image_ori = None self.dataset = opt.dataset if self.dataset == "nuscenes": self.tracker = {} for class_name in NUSCENES_TRACKING_NAMES: self.tracker[class_name] = Tracker(opt, self.model) else: self.tracker = Tracker(opt, self.model) self.debugger = Debugger(opt=opt, dataset=self.trained_dataset) self.img_height = 100 self.img_width = 100 def run(self, image_or_path_or_tensor, meta={}, image_info=None): load_time, pre_time, net_time, dec_time, post_time = 0, 0, 0, 0, 0 merge_time, track_time, tot_time, display_time = 0, 0, 0, 0 self.debugger.clear() start_time = time.time() # read image pre_processed = False if isinstance(image_or_path_or_tensor, np.ndarray): image = image_or_path_or_tensor elif type(image_or_path_or_tensor) == type(""): image = cv2.imread(image_or_path_or_tensor) else: image = image_or_path_or_tensor["image"][0].numpy() pre_processed_images = image_or_path_or_tensor pre_processed = True loaded_time = time.time() load_time += loaded_time - start_time detections = [] # for multi-scale testing for scale in self.opt.test_scales: scale_start_time = time.time() if not pre_processed: # not prefetch testing or demo images, meta = self.pre_process(image, scale, meta) else: # prefetch testing images = pre_processed_images["images"][scale][0] meta = pre_processed_images["meta"][scale] meta = {k: v.numpy()[0] for k, v in meta.items()} if "pre_dets" in pre_processed_images["meta"]: meta["pre_dets"] = pre_processed_images["meta"]["pre_dets"] if "cur_dets" in pre_processed_images["meta"]: meta["cur_dets"] = pre_processed_images["meta"]["cur_dets"] images = images.to(self.opt.device, non_blocking=self.opt.non_block_test) # initializing tracker pre_hms, pre_inds = None, None pre_process_time = time.time() pre_time += pre_process_time - scale_start_time # run the network # output: the output feature maps, only used for visualizing # dets: output tensors after extracting peaks output, dets, forward_time, FeatureMaps = self.process( images, self.pre_images, pre_hms, pre_inds, return_time=True) net_time += forward_time - pre_process_time decode_time = time.time() dec_time += decode_time - forward_time # convert the cropped and 4x downsampled output coordinate system # back to the input image coordinate system result = self.post_process(dets, meta, scale) post_process_time = time.time() post_time += post_process_time - decode_time detections.append(result) if self.opt.debug >= 2: self.debug( self.debugger, images, result, output, scale, pre_images=self.pre_images if not self.opt.no_pre_img else None, pre_hms=pre_hms, ) # merge multi-scale testing results results = self.merge_outputs(detections) torch.cuda.synchronize() end_time = time.time() merge_time += end_time - post_process_time # public detection mode in MOT challenge if self.opt.public_det: results = (pre_processed_images["meta"]["cur_dets"] if self.opt.public_det else None) if self.dataset == "nuscenes": trans_matrix = np.array(image_info["trans_matrix"], np.float32) results_by_class = {} ddd_boxes_by_class = {} depths_by_class = {} ddd_boxes_by_class2 = {} ddd_org_boxes_by_class = {} ddd_box_submission1 = {} ddd_box_submission2 = {} for class_name in NUSCENES_TRACKING_NAMES: results_by_class[class_name] = [] ddd_boxes_by_class2[class_name] = [] ddd_boxes_by_class[class_name] = [] depths_by_class[class_name] = [] ddd_org_boxes_by_class[class_name] = [] ddd_box_submission1[class_name] = [] ddd_box_submission2[class_name] = [] for det in results: cls_id = int(det["class"]) class_name = nuscenes_class_name[cls_id - 1] if class_name not in NUSCENES_TRACKING_NAMES: continue if det["score"] < 0.3: continue if class_name == "pedestrian" and det["score"] < 0.35: continue results_by_class[class_name].append(det["bbox"].tolist() + [det["score"]]) size = [ float(det["dim"][1]), float(det["dim"][2]), float(det["dim"][0]), ] rot_cam = Quaternion(axis=[0, 1, 0], angle=det["rot_y"]) translation_submission1 = np.dot( trans_matrix, np.array( [ det["loc"][0], det["loc"][1] - size[2], det["loc"][2], 1 ], np.float32, ), ).copy() loc = np.array([det["loc"][0], det["loc"][1], det["loc"][2]], np.float32) depths_by_class[class_name].append([float(det["loc"][2]) ].copy()) trans = [det["loc"][0], det["loc"][1], det["loc"][2]] ddd_org_boxes_by_class[class_name].append([ float(det["dim"][0]), float(det["dim"][1]), float(det["dim"][2]) ] + trans + [det["rot_y"]]) box = Box(loc, size, rot_cam, name="2", token="1") box.translate(np.array([0, -box.wlh[2] / 2, 0])) box.rotate(Quaternion(image_info["cs_record_rot"])) box.translate(np.array(image_info["cs_record_trans"])) box.rotate(Quaternion(image_info["pose_record_rot"])) box.translate(np.array(image_info["pose_record_trans"])) rotation = box.orientation rotation = [ float(rotation.w), float(rotation.x), float(rotation.y), float(rotation.z), ] ddd_box_submission1[class_name].append([ float(translation_submission1[0]), float(translation_submission1[1]), float(translation_submission1[2]), ].copy() + size.copy() + rotation.copy()) q = Quaternion(rotation) angle = q.angle if q.axis[2] > 0 else -q.angle ddd_boxes_by_class[class_name].append([ size[2], size[0], size[1], box.center[0], box.center[1], box.center[2], angle, ].copy()) online_targets = [] for class_name in NUSCENES_TRACKING_NAMES: if len(results_by_class[class_name]) > 0 and NMS: boxess = torch.from_numpy( np.array(results_by_class[class_name])[:, :4]) scoress = torch.from_numpy( np.array(results_by_class[class_name])[:, -1]) if class_name == "bus" or class_name == "truck": ovrlp = 0.7 else: ovrlp = 0.8 keep, count = nms(boxess, scoress, overlap=ovrlp) keep = keep.data.numpy().tolist() keep = sorted(set(keep)) results_by_class[class_name] = np.array( results_by_class[class_name])[keep] ddd_boxes_by_class[class_name] = np.array( ddd_boxes_by_class[class_name])[keep] depths_by_class[class_name] = np.array( depths_by_class[class_name])[keep] ddd_org_boxes_by_class[class_name] = np.array( ddd_org_boxes_by_class[class_name])[keep] ddd_box_submission1[class_name] = np.array( ddd_box_submission1[class_name])[keep] online_targets += self.tracker[class_name].update( results_by_class[class_name], FeatureMaps, ddd_boxes=ddd_boxes_by_class[class_name], depths_by_class=depths_by_class[class_name], ddd_org_boxes=ddd_org_boxes_by_class[class_name], submission=ddd_box_submission1[class_name], classe=class_name, ) else: online_targets = self.tracker.update(results, FeatureMaps) return online_targets def _transform_scale(self, image, scale=1): """ Prepare input image in different testing modes. Currently support: fix short size/ center crop to a fixed size/ keep original resolution but pad to a multiplication of 32 """ height, width = image.shape[0:2] new_height = int(height * scale) new_width = int(width * scale) if self.opt.fix_short > 0: if height < width: inp_height = self.opt.fix_short inp_width = (int(width / height * self.opt.fix_short) + 63) // 64 * 64 else: inp_height = (int(height / width * self.opt.fix_short) + 63) // 64 * 64 inp_width = self.opt.fix_short c = np.array([width / 2, height / 2], dtype=np.float32) s = np.array([width, height], dtype=np.float32) elif self.opt.fix_res: inp_height, inp_width = self.opt.input_h, self.opt.input_w c = np.array([new_width / 2.0, new_height / 2.0], dtype=np.float32) s = max(height, width) * 1.0 # s = np.array([inp_width, inp_height], dtype=np.float32) else: inp_height = (new_height | self.opt.pad) + 1 inp_width = (new_width | self.opt.pad) + 1 c = np.array([new_width // 2, new_height // 2], dtype=np.float32) s = np.array([inp_width, inp_height], dtype=np.float32) resized_image = cv2.resize(image, (new_width, new_height)) return resized_image, c, s, inp_width, inp_height, height, width def pre_process(self, image, scale, input_meta={}): """ Crop, resize, and normalize image. Gather meta data for post processing and tracking. """ resized_image, c, s, inp_width, inp_height, height, width = self._transform_scale( image) trans_input = get_affine_transform(c, s, 0, [inp_width, inp_height]) out_height = inp_height // self.opt.down_ratio out_width = inp_width // self.opt.down_ratio trans_output = get_affine_transform(c, s, 0, [out_width, out_height]) inp_image = cv2.warpAffine(resized_image, trans_input, (inp_width, inp_height), flags=cv2.INTER_LINEAR) inp_image = ((inp_image / 255.0 - self.mean) / self.std).astype( np.float32) images = inp_image.transpose(2, 0, 1).reshape(1, 3, inp_height, inp_width) if self.opt.flip_test: images = np.concatenate((images, images[:, :, :, ::-1]), axis=0) images = torch.from_numpy(images) meta = { "calib": np.array(input_meta["calib"], dtype=np.float32) if "calib" in input_meta else self._get_default_calib(width, height) } meta.update({ "c": c, "s": s, "height": height, "width": width, "out_height": out_height, "out_width": out_width, "inp_height": inp_height, "inp_width": inp_width, "trans_input": trans_input, "trans_output": trans_output, }) if "pre_dets" in input_meta: meta["pre_dets"] = input_meta["pre_dets"] if "cur_dets" in input_meta: meta["cur_dets"] = input_meta["cur_dets"] return images, meta def _trans_bbox(self, bbox, trans, width, height): """ Transform bounding boxes according to image crop. """ bbox = np.array(copy.deepcopy(bbox), dtype=np.float32) bbox[:2] = affine_transform(bbox[:2], trans) bbox[2:] = affine_transform(bbox[2:], trans) bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, width - 1) bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, height - 1) return bbox def _get_additional_inputs(self, dets, meta, with_hm=True): """ Render input heatmap from previous trackings. """ trans_input, trans_output = meta["trans_input"], meta["trans_output"] inp_width, inp_height = meta["inp_width"], meta["inp_height"] out_width, out_height = meta["out_width"], meta["out_height"] input_hm = np.zeros((1, inp_height, inp_width), dtype=np.float32) output_inds = [] for det in dets: if det["score"] < self.opt.pre_thresh or det["active"] == 0: continue bbox = self._trans_bbox(det["bbox"], trans_input, inp_width, inp_height) bbox_out = self._trans_bbox(det["bbox"], trans_output, out_width, out_height) h, w = bbox[3] - bbox[1], bbox[2] - bbox[0] if h > 0 and w > 0: radius = gaussian_radius((math.ceil(h), math.ceil(w))) radius = max(0, int(radius)) ct = np.array([(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2], dtype=np.float32) ct_int = ct.astype(np.int32) if with_hm: draw_umich_gaussian(input_hm[0], ct_int, radius) ct_out = np.array( [(bbox_out[0] + bbox_out[2]) / 2, (bbox_out[1] + bbox_out[3]) / 2], dtype=np.int32, ) output_inds.append(ct_out[1] * out_width + ct_out[0]) if with_hm: input_hm = input_hm[np.newaxis] if self.opt.flip_test: input_hm = np.concatenate((input_hm, input_hm[:, :, :, ::-1]), axis=0) input_hm = torch.from_numpy(input_hm).to(self.opt.device) output_inds = np.array(output_inds, np.int64).reshape(1, -1) output_inds = torch.from_numpy(output_inds).to(self.opt.device) return input_hm, output_inds def _get_default_calib(self, width, height): calib = np.array([ [self.rest_focal_length, 0, width / 2, 0], [0, self.rest_focal_length, height / 2, 0], [0, 0, 1, 0], ]) return calib def _sigmoid_output(self, output): if "hm" in output: output["hm"] = output["hm"].sigmoid_() if "hm_hp" in output: output["hm_hp"] = output["hm_hp"].sigmoid_() if "dep" in output: output["dep"] = 1.0 / (output["dep"].sigmoid() + 1e-6) - 1.0 output["dep"] *= self.opt.depth_scale return output def _flip_output(self, output): average_flips = ["hm", "wh", "dep", "dim"] neg_average_flips = ["amodel_offset"] single_flips = [ "ltrb", "nuscenes_att", "velocity", "ltrb_amodal", "reg", "hp_offset", "rot", "tracking", "pre_hm", ] for head in output: if head in average_flips: output[head] = (output[head][0:1] + flip_tensor(output[head][1:2])) / 2 if head in neg_average_flips: flipped_tensor = flip_tensor(output[head][1:2]) flipped_tensor[:, 0::2] *= -1 output[head] = (output[head][0:1] + flipped_tensor) / 2 if head in single_flips: output[head] = output[head][0:1] if head == "hps": output["hps"] = (output["hps"][0:1] + flip_lr_off( output["hps"][1:2], self.flip_idx)) / 2 if head == "hm_hp": output["hm_hp"] = (output["hm_hp"][0:1] + flip_lr( output["hm_hp"][1:2], self.flip_idx)) / 2 return output def process(self, images, pre_images=None, pre_hms=None, pre_inds=None, return_time=False): with torch.no_grad(): torch.cuda.synchronize() output, FeatureMaps = self.model(images, pre_images, pre_hms) output = output[-1] output = self._sigmoid_output(output) output.update({"pre_inds": pre_inds}) if self.opt.flip_test: output = self._flip_output(output) torch.cuda.synchronize() forward_time = time.time() dets = generic_decode(output, K=self.opt.K, opt=self.opt) torch.cuda.synchronize() for k in dets: dets[k] = dets[k].detach().cpu().numpy() if return_time: return output, dets, forward_time, FeatureMaps else: return output, dets, FeatureMaps def post_process(self, dets, meta, scale=1): dets = generic_post_process( self.opt, dets, [meta["c"]], [meta["s"]], meta["out_height"], meta["out_width"], self.opt.num_classes, [meta["calib"]], meta["height"], meta["width"], ) self.this_calib = meta["calib"] if scale != 1: for i in range(len(dets[0])): for k in ["bbox", "hps"]: if k in dets[0][i]: dets[0][i][k] = (np.array(dets[0][i][k], np.float32) / scale).tolist() return dets[0] def merge_outputs(self, detections): assert len(self.opt.test_scales) == 1, "multi_scale not supported!" results = [] for i in range(len(detections[0])): if detections[0][i]["score"] > self.opt.out_thresh: results.append(detections[0][i]) return results def debug(self, debugger, images, dets, output, scale=1, pre_images=None, pre_hms=None): img = images[0].detach().cpu().numpy().transpose(1, 2, 0) img = np.clip(((img * self.std + self.mean) * 255.0), 0, 255).astype(np.uint8) pred = debugger.gen_colormap(output["hm"][0].detach().cpu().numpy()) debugger.add_blend_img(img, pred, "pred_hm") if "hm_hp" in output: pred = debugger.gen_colormap_hp( output["hm_hp"][0].detach().cpu().numpy()) debugger.add_blend_img(img, pred, "pred_hmhp") if pre_images is not None: pre_img = pre_images[0].detach().cpu().numpy().transpose(1, 2, 0) pre_img = np.clip(((pre_img * self.std + self.mean) * 255.0), 0, 255).astype(np.uint8) debugger.add_img(pre_img, "pre_img") if pre_hms is not None: pre_hm = debugger.gen_colormap( pre_hms[0].detach().cpu().numpy()) debugger.add_blend_img(pre_img, pre_hm, "pre_hm") def show_results(self, debugger, image, results): debugger.add_img(image, img_id="generic") if self.opt.tracking: debugger.add_img( self.pre_image_ori if self.pre_image_ori is not None else image, img_id="previous", ) self.pre_image_ori = image for j in range(len(results)): if results[j]["score"] > self.opt.vis_thresh: if "active" in results[j] and results[j]["active"] == 0: continue item = results[j] if "bbox" in item: sc = (item["score"] if self.opt.demo == "" or not ("tracking_id" in item) else item["tracking_id"]) sc = item[ "tracking_id"] if self.opt.show_track_color else sc debugger.add_coco_bbox(item["bbox"], item["class"] - 1, sc, img_id="generic") if "tracking" in item: debugger.add_arrow(item["ct"], item["tracking"], img_id="generic") tracking_id = item[ "tracking_id"] if "tracking_id" in item else -1 if ("tracking_id" in item and self.opt.demo == "" and not self.opt.show_track_color): debugger.add_tracking_id(item["ct"], item["tracking_id"], img_id="generic") if (item["class"] in [1, 2]) and "hps" in item: debugger.add_coco_hp(item["hps"], tracking_id=tracking_id, img_id="generic") if (len(results) > 0 and "dep" in results[0] and "alpha" in results[0] and "dim" in results[0]): debugger.add_3d_detection( image if not self.opt.qualitative else cv2.resize( debugger.imgs["pred_hm"], (image.shape[1], image.shape[0])), False, results, self.this_calib, vis_thresh=self.opt.vis_thresh, img_id="ddd_pred", ) debugger.add_bird_view( results, vis_thresh=self.opt.vis_thresh, img_id="bird_pred", cnt=self.cnt, ) if self.opt.show_track_color and self.opt.debug == 4: del debugger.imgs["generic"], debugger.imgs["bird_pred"] def reset_tracking(self, opt): if self.dataset == "nuscenes": self.tracker = {} for class_name in NUSCENES_TRACKING_NAMES: self.tracker[class_name] = Tracker(opt, self.model, h=self.img_height, w=self.img_width) else: self.tracker = Tracker(opt, self.model, h=self.img_height, w=self.img_width) self.pre_images = None self.pre_image_ori = None def update_public_detections(self, detections_file): self.det_file = pd.read_csv(detections_file, header=None, sep=" ") self.det_group = self.det_file.groupby(0) self.det_group_keys = self.det_group.indices.keys()