def main(): cfg.device = torch.device('cuda') torch.backends.cudnn.benchmark = False max_per_image = 100 num_classes = 80 if cfg.dataset == 'coco' else 4 dictionary = np.load(cfg.dictionary_file) colors = COCO_COLORS if cfg.dataset == 'coco' else DETRAC_COLORS names = COCO_NAMES if cfg.dataset == 'coco' else DETRAC_NAMES for j in range(len(names)): col_ = [c * 255 for c in colors[j]] colors[j] = tuple(col_) print('Creating model and recover from checkpoint ...') if 'hourglass' in cfg.arch: model = exkp(n=5, nstack=2, dims=[256, 256, 384, 384, 384, 512], modules=[2, 2, 2, 2, 2, 4], num_classes=num_classes) else: model = get_pose_net(num_layers=int(cfg.arch.split('_')[-1]), num_classes=80) # raise NotImplementedError model = load_demo_model(model, cfg.ckpt_dir) model = model.to(cfg.device) model.eval() # Loading COCO validation images annotation_file = '{}/annotations/instances_{}.json'.format( cfg.data_dir, cfg.data_type) coco = COCO(annotation_file) # Load all annotations cats = coco.loadCats(coco.getCatIds()) nms = [cat['name'] for cat in cats] catIds = coco.getCatIds(catNms=nms) # imgIds = coco.getImgIds(catIds=catIds) imgIds = coco.getImgIds() # annIds = coco.getAnnIds(catIds=catIds) # all_anns = coco.loadAnns(ids=annIds) # print(len(imgIds), imgIds) for id in imgIds: annt_ids = coco.getAnnIds(imgIds=[id]) annotations_per_img = coco.loadAnns(ids=annt_ids) # print('All annots: ', len(annotations_per_img), annotations_per_img) img = coco.loadImgs(id)[0] image_path = '%s/images/%s/%s' % (cfg.data_dir, cfg.data_type, img['file_name']) w_img = int(img['width']) h_img = int(img['height']) if w_img < 1 or h_img < 1: continue img_original = cv2.imread(image_path) img_connect = cv2.imread(image_path) img_recon = cv2.imread(image_path) print('Image id: ', id) for annt in annotations_per_img: if annt['iscrowd'] == 1 or type(annt['segmentation']) != list: continue polygons = get_connected_polygon_using_mask( annt['segmentation'], (h_img, w_img), n_vertices=cfg.num_vertices, closing_max_kernel=60) gt_bbox = annt['bbox'] gt_x1, gt_y1, gt_w, gt_h = gt_bbox contour = np.array(polygons).reshape((-1, 2)) # Downsample the contour to fix number of vertices if len(contour) > cfg.num_vertices: resampled_contour = resample(contour, num=cfg.num_vertices) else: resampled_contour = turning_angle_resample( contour, cfg.num_vertices) resampled_contour[:, 0] = np.clip(resampled_contour[:, 0], gt_x1, gt_x1 + gt_w) resampled_contour[:, 1] = np.clip(resampled_contour[:, 1], gt_y1, gt_y1 + gt_h) clockwise_flag = check_clockwise_polygon(resampled_contour) if not clockwise_flag: fixed_contour = np.flip(resampled_contour, axis=0) else: fixed_contour = resampled_contour.copy() # Indexing from the left-most vertex, argmin x-axis idx = np.argmin(fixed_contour[:, 0]) indexed_shape = np.concatenate( (fixed_contour[idx:, :], fixed_contour[:idx, :]), axis=0) x1, y1, x2, y2 = gt_x1, gt_y1, gt_x1 + gt_w, gt_y1 + gt_h # bbox_width, bbox_height = x2 - x1, y2 - y1 # bbox = [x1, y1, bbox_width, bbox_height] # bbox_center = np.array([(x1 + x2) / 2., (y1 + y2) / 2.]) bbox_center = np.mean(indexed_shape, axis=0) centered_shape = indexed_shape - bbox_center # visualize resampled points with multiple parts in image side by side for cnt in range(len(annt['segmentation'])): polys = np.array(annt['segmentation'][cnt]).reshape((-1, 2)) cv2.polylines(img_original, [polys.astype(np.int32)], True, (10, 10, 255), thickness=2) # cv2.drawContours(img_original, [polys.astype(np.int32)], contourIdx=-1, color=(10, 10, 255), thickness=-1) cv2.polylines(img_connect, [indexed_shape.astype(np.int32)], True, (10, 10, 255), thickness=2) # cv2.drawContours(img_connect, [indexed_shape.astype(np.int32)], contourIdx=-1, color=(10, 10, 255), thickness=-1) learned_val_codes, _ = fast_ista(centered_shape.reshape((1, -1)), dictionary, lmbda=0.1, max_iter=60) recon_contour = np.matmul(learned_val_codes, dictionary).reshape( (-1, 2)) recon_contour = recon_contour + bbox_center cv2.polylines(img_recon, [recon_contour.astype(np.int32)], True, (10, 10, 255), thickness=2) # cv2.drawContours(img_recon, [recon_contour.astype(np.int32)], contourIdx=-1, color=(10, 10, 255), thickness=-1) # plot gt mean and std # image = cv2.imread(image_path) # # cv2.ellipse(image, center=(int(contour_mean[0]), int(contour_mean[1])), # # axes=(int(contour_std[0]), int(contour_std[1])), # # angle=0, startAngle=0, endAngle=360, color=(0, 255, 0), # # thickness=2) # cv2.rectangle(image, pt1=(int(contour_mean[0] - contour_std[0] / 2.), int(contour_mean[1] - contour_std[1] / 2.)), # pt2=(int(contour_mean[0] + contour_std[0] / 2.), int(contour_mean[1] + contour_std[1] / 2.)), # color=(0, 255, 0), thickness=2) # cv2.polylines(image, [fixed_contour.astype(np.int32)], True, (0, 0, 255)) # cv2.rectangle(image, pt1=(int(min(fixed_contour[:, 0])), int(min(fixed_contour[:, 1]))), # pt2=(int(max(fixed_contour[:, 0])), int(max(fixed_contour[:, 1]))), # color=(255, 0, 0), thickness=2) # cv2.imshow('GT segments', image) # if cv2.waitKey() & 0xFF == ord('q'): # break image = cv2.imread(image_path) original_image = image.copy() height, width = image.shape[0:2] padding = 127 if 'hourglass' in cfg.arch else 31 imgs = {} for scale in cfg.test_scales: new_height = int(height * scale) new_width = int(width * scale) if cfg.img_size > 0: img_height, img_width = cfg.img_size, cfg.img_size center = np.array([new_width / 2., new_height / 2.], dtype=np.float32) scaled_size = max(height, width) * 1.0 scaled_size = np.array([scaled_size, scaled_size], dtype=np.float32) else: img_height = (new_height | padding) + 1 img_width = (new_width | padding) + 1 center = np.array([new_width // 2, new_height // 2], dtype=np.float32) scaled_size = np.array([img_width, img_height], dtype=np.float32) img = cv2.resize(image, (new_width, new_height)) trans_img = get_affine_transform(center, scaled_size, 0, [img_width, img_height]) img = cv2.warpAffine(img, trans_img, (img_width, img_height)) img = img.astype(np.float32) / 255. img -= np.array( COCO_MEAN if cfg.dataset == 'coco' else DETRAC_MEAN, dtype=np.float32)[None, None, :] img /= np.array(COCO_STD if cfg.dataset == 'coco' else DETRAC_STD, dtype=np.float32)[None, None, :] img = img.transpose( 2, 0, 1)[None, :, :, :] # from [H, W, C] to [1, C, H, W] # if cfg.test_flip: # img = np.concatenate((img, img[:, :, :, ::-1].copy()), axis=0) imgs[scale] = { 'image': torch.from_numpy(img).float(), 'center': np.array(center), 'scale': np.array(scaled_size), 'fmap_h': np.array(img_height // 4), 'fmap_w': np.array(img_width // 4) } with torch.no_grad(): segmentations = [] predicted_codes = [] start_time = time.time() for scale in imgs: imgs[scale]['image'] = imgs[scale]['image'].to(cfg.device) output = model(imgs[scale]['image'])[-1] # segms, codes_ = ctsegm_scaled_decode_debug(*output, torch.from_numpy(dictionary.astype(np.float32)).to(cfg.device), # K=cfg.test_topk) segms = ctsegm_code_n_offset_decode( *output, torch.from_numpy(dictionary.astype(np.float32)).to( cfg.device), K=cfg.test_topk) segms = segms.detach().cpu().numpy().reshape( 1, -1, segms.shape[2])[0] # codes_ = codes_.detach().cpu().numpy().reshape(1, -1, codes_.shape[2])[0] top_preds = {} code_preds = {} for j in range(cfg.num_vertices): segms[:, 2 * j:2 * j + 2] = transform_preds( segms[:, 2 * j:2 * j + 2], imgs[scale]['center'], imgs[scale]['scale'], (imgs[scale]['fmap_w'], imgs[scale]['fmap_h'])) segms[:, cfg.num_vertices * 2:cfg.num_vertices * 2 + 2] = transform_preds( segms[:, cfg.num_vertices * 2:cfg.num_vertices * 2 + 2], imgs[scale]['center'], imgs[scale]['scale'], (imgs[scale]['fmap_w'], imgs[scale]['fmap_h'])) segms[:, cfg.num_vertices * 2 + 2:cfg.num_vertices * 2 + 4] = transform_preds( segms[:, cfg.num_vertices * 2 + 2:cfg.num_vertices * 2 + 4], imgs[scale]['center'], imgs[scale]['scale'], (imgs[scale]['fmap_w'], imgs[scale]['fmap_h'])) clses = segms[:, -1] for j in range(num_classes): inds = (clses == j) top_preds[j + 1] = segms[inds, :cfg.num_vertices * 2 + 5].astype(np.float32) top_preds[j + 1][:, :cfg.num_vertices * 2 + 4] /= scale # code_preds[j + 1] = codes_[inds, :] segmentations.append(top_preds) predicted_codes.append(code_preds) segms_and_scores = { j: np.concatenate([d[j] for d in segmentations], axis=0) for j in range(1, num_classes + 1) } # a Dict label: segments # codes_and_scores = {j: np.concatenate([d[j] for d in predicted_codes], axis=0) # for j in range(1, num_classes + 1)} # a Dict label: segments scores = np.hstack([ segms_and_scores[j][:, cfg.num_vertices * 2 + 4] for j in range(1, num_classes + 1) ]) if len(scores) > max_per_image: kth = len(scores) - max_per_image thresh = np.partition(scores, kth)[kth] for j in range(1, num_classes + 1): keep_inds = (segms_and_scores[j][:, cfg.num_vertices * 2 + 4] >= thresh) segms_and_scores[j] = segms_and_scores[j][keep_inds] # codes_and_scores[j] = codes_and_scores[j][keep_inds] # Use opencv functions to output a video output_image = original_image blend_mask = np.zeros(shape=output_image.shape, dtype=np.uint8) # print(blend_mask.shape) for lab in segms_and_scores: for idx in range(len(segms_and_scores[lab])): res = segms_and_scores[lab][idx] # c_ = codes_and_scores[lab][idx] # for res in segms_and_scores[lab]: contour, bbox, score = res[:-5], res[-5:-1], res[-1] bbox[0] = np.clip(bbox[0], 0, w_img) bbox[1] = np.clip(bbox[1], 0, h_img) bbox[2] = np.clip(bbox[2], 0, w_img) bbox[3] = np.clip(bbox[3], 0, h_img) if score > cfg.detect_thres: text = names[lab] # + ' %.2f' % score # label_size = cv2.getTextSize(text, cv2.FONT_HERSHEY_COMPLEX, thickness=2, fontScale=0.5) polygon = contour.reshape((-1, 2)) # print('Shape: Poly -- ', polygon.shape) # print(polygon) polygon[:, 0] = np.clip(polygon[:, 0], 0, w_img - 1) polygon[:, 1] = np.clip(polygon[:, 1], 0, h_img - 1) # use bb tools to draw predictions color = random.choice(COLOR_WORLD) bb.add(output_image, bbox[0], bbox[1], bbox[2], bbox[3], text, color) cv2.polylines(output_image, [polygon.astype(np.int32)], True, RGB_DICT[color], thickness=1) cv2.drawContours(blend_mask, [polygon.astype(np.int32)], contourIdx=-1, color=RGB_DICT[color], thickness=-1) # color = (random.randint(0, 255), random.randint(0, 255), random.randint(0, 255)) # contour_mean = np.mean(polygon, axis=0) # contour_std = np.std(polygon, axis=0) # center_x, center_y = np.mean(polygon, axis=0).astype(np.int32) # text_location = [bbox[0] + 1, bbox[1] + 1, # bbox[1] + label_size[0][0] + 1, # bbox[0] + label_size[0][1] + 1] # cv2.rectangle(output_image, pt1=(int(bbox[0]), int(bbox[1])), # pt2=(int(bbox[2]), int(bbox[3])), # color=color, thickness=1) # cv2.rectangle(output_image, pt1=(int(np.min(polygon[:, 0])), int(np.min(polygon[:, 1]))), # pt2=(int(np.max(polygon[:, 0])), int(np.max(polygon[:, 1]))), # color=(0, 255, 0), thickness=1) # cv2.polylines(output_image, [polygon.astype(np.int32)], True, color, thickness=2) # cv2.putText(output_image, text, org=(int(text_location[0]), int(text_location[3])), # fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=2, fontScale=0.5, # color=(255, 0, 0)) # cv2.putText(output_image, text, org=(int(bbox[0]), int(bbox[1])), # fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=1, fontScale=0.5, # color=color) # show the histgram for predicted codes # fig = plt.figure() # plt.plot(np.arange(cfg.n_codes), c_.reshape((-1,)), color='green', # marker='o', linestyle='dashed', linewidth=2, markersize=6) # plt.ylabel('Value of each coefficient') # plt.xlabel('All predicted {} coefficients'.format(cfg.n_codes)) # plt.title('Distribution of the predicted coefficients for {}'.format(text)) # plt.show() value = [255, 255, 255] dst_img = cv2.addWeighted(output_image, 0.5, blend_mask, 0.5, 0) dst_img[blend_mask == 0] = output_image[blend_mask == 0] img_original = cv2.copyMakeBorder(img_original, 0, 0, 0, 10, cv2.BORDER_CONSTANT, None, value) img_connect = cv2.copyMakeBorder(img_connect, 0, 0, 10, 10, cv2.BORDER_CONSTANT, None, value) img_recon = cv2.copyMakeBorder(img_recon, 0, 0, 10, 10, cv2.BORDER_CONSTANT, None, value) dst_img = cv2.copyMakeBorder(dst_img, 0, 0, 10, 0, cv2.BORDER_CONSTANT, None, value) im_cat = np.concatenate( (img_original, img_connect, img_recon, dst_img), axis=1) # im_cat = np.concatenate((img_original, img_connect, img_recon), axis=1) cv2.imshow('GT:Resample:Recons:Predict', im_cat) if cv2.waitKey() & 0xFF == ord('q'): break
def __getitem__(self, index): img_id = self.images[index] img_path = os.path.join( self.img_dir, self.coco.loadImgs(ids=[img_id])[0]['file_name']) ann_ids = self.coco.getAnnIds(imgIds=[img_id]) annotations = self.coco.loadAnns(ids=ann_ids) img = self.coco.loadImgs(ids=[img_id])[0] w_img = int(img['width']) h_img = int(img['height']) labels = [] bboxes = [] shapes = [] for anno in annotations: if anno['iscrowd'] == 1: # Excludes crowd objects continue polygons = get_connected_polygon_using_mask( anno['segmentation'], (h_img, w_img), n_vertices=self.n_vertices, closing_max_kernel=50) gt_x1, gt_y1, gt_w, gt_h = anno['bbox'] contour = np.array(polygons).reshape((-1, 2)) # Downsample the contour to fix number of vertices if len(contour) > self.n_vertices: fixed_contour = resample(contour, num=self.n_vertices) else: fixed_contour = turning_angle_resample(contour, self.n_vertices) fixed_contour[:, 0] = np.clip(fixed_contour[:, 0], gt_x1, gt_x1 + gt_w) fixed_contour[:, 1] = np.clip(fixed_contour[:, 1], gt_y1, gt_y1 + gt_h) contour_std = np.sqrt(np.sum(np.std(fixed_contour, axis=0)**2)) if contour_std < 1e-6 or contour_std == np.inf or contour_std == np.nan: # invalid shapes continue updated_bbox = [ np.min(fixed_contour[:, 0]), np.min(fixed_contour[:, 1]), np.max(fixed_contour[:, 0]), np.max(fixed_contour[:, 1]) ] shapes.append(np.ndarray.flatten(fixed_contour).tolist()) labels.append(self.cat_ids[anno['category_id']]) bboxes.append(updated_bbox) labels = np.array(labels) bboxes = np.array(bboxes, dtype=np.float32) shapes = np.array(shapes, dtype=np.float32) if len(bboxes) == 0: bboxes = np.array([[0., 0., 0., 0.]], dtype=np.float32) labels = np.array([[0]]) shapes = np.zeros((1, self.n_vertices * 2), dtype=np.float32) # bboxes[:, 2:] += bboxes[:, :2] # xywh to xyxy img = cv2.imread(img_path) height, width = img.shape[0], img.shape[1] center = np.array([width / 2., height / 2.], dtype=np.float32) # center of image scale = max(height, width) * 1.0 flipped = False if self.split == 'train': scale = scale * np.random.choice(self.rand_scales) w_border = get_border(160, width) h_border = get_border(160, height) center[0] = np.random.randint(low=w_border, high=width - w_border) center[1] = np.random.randint(low=h_border, high=height - h_border) if np.random.random() < 0.5: flipped = True img = img[:, ::-1, :] center[0] = width - center[0] - 1 trans_img = get_affine_transform( center, scale, 0, [self.img_size['w'], self.img_size['h']]) img = cv2.warpAffine(img, trans_img, (self.img_size['w'], self.img_size['h'])) # -----------------------------------debug--------------------------------- # image_show = img.copy() # for bbox, label in zip(bboxes, labels): # if flipped: # bbox[[0, 2]] = width - bbox[[2, 0]] - 1 # bbox[:2] = affine_transform(bbox[:2], trans_img) # bbox[2:] = affine_transform(bbox[2:], trans_img) # bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, self.img_size['w'] - 1) # bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, self.img_size['h'] - 1) # cv2.rectangle(image_show, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])), (255, 0, 0), 2) # cv2.putText(image_show, self.class_name[label + 1], (int(bbox[0]), int(bbox[1])), # cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 1) # cv2.imshow('img', image_show) # cv2.waitKey() # -----------------------------------debug--------------------------------- img = img.astype(np.float32) / 255. if self.split == 'train': color_aug(self.data_rng, img, self.eig_val, self.eig_vec) img -= self.mean img /= self.std img = img.transpose(2, 0, 1) # from [H, W, C] to [C, H, W] trans_fmap = get_affine_transform( center, scale, 0, [self.fmap_size['w'], self.fmap_size['h']]) hmap = np.zeros( (self.num_classes, self.fmap_size['h'], self.fmap_size['w']), dtype=np.float32) # heatmap w_h_ = np.zeros((self.max_objs, 2), dtype=np.float32) # width and height of bboxes shapes_ = np.zeros((self.max_objs, self.n_vertices * 2), dtype=np.float32) # gt amodal segmentation polygons center_offsets = np.zeros( (self.max_objs, 2), dtype=np.float32) # gt mass centers to bbox center codes_ = np.zeros((self.max_objs, self.n_codes), dtype=np.float32) contour_std_ = np.zeros( (self.max_objs, 1), dtype=np.float32) # keep track of codes that is activated regs = np.zeros( (self.max_objs, 2), dtype=np.float32) # regression for offsets of shape center inds = np.zeros((self.max_objs, ), dtype=np.int64) ind_masks = np.zeros((self.max_objs, ), dtype=np.uint8) for k, (bbox, label, shape) in enumerate(zip(bboxes, labels, shapes)): if flipped: bbox[[0, 2]] = width - bbox[[2, 0]] - 1 # Flip the contour x-axis for m in range(self.n_vertices): shape[2 * m] = width - shape[2 * m] - 1 bbox[:2] = affine_transform(bbox[:2], trans_fmap) bbox[2:] = affine_transform(bbox[2:], trans_fmap) bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, self.fmap_size['w'] - 1) bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, self.fmap_size['h'] - 1) h, w = bbox[3] - bbox[1], bbox[2] - bbox[0] # generate gt shape mean and std from contours for m in range(self.n_vertices ): # apply scale and crop transform to shapes shape[2 * m:2 * m + 2] = affine_transform( shape[2 * m:2 * m + 2], trans_fmap) shape_clipped = np.reshape(shape, (self.n_vertices, 2)) shape_clipped[:, 0] = np.clip(shape_clipped[:, 0], 0, self.fmap_size['w'] - 1) shape_clipped[:, 1] = np.clip(shape_clipped[:, 1], 0, self.fmap_size['h'] - 1) clockwise_flag = check_clockwise_polygon(shape_clipped) if not clockwise_flag: fixed_contour = np.flip(shape_clipped, axis=0) else: fixed_contour = shape_clipped.copy() # Indexing from the left-most vertex, argmin x-axis idx = np.argmin(fixed_contour[:, 0]) indexed_shape = np.concatenate( (fixed_contour[idx:, :], fixed_contour[:idx, :]), axis=0) mass_center = np.mean(indexed_shape, axis=0) contour_std = np.std(indexed_shape, axis=0) + 1e-4 if h < 1e-6 or w < 1e-6: # remove small bboxes continue # centered_shape = indexed_shape - mass_center norm_shape = (indexed_shape - mass_center) / np.sqrt( np.sum(contour_std**2)) if h > 0 and w > 0: obj_c = np.array([(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2], dtype=np.float32) # obj_c = mass_center obj_c_int = obj_c.astype(np.int32) radius = max( 0, int( gaussian_radius((math.ceil(h), math.ceil(w)), self.gaussian_iou))) draw_umich_gaussian(hmap[label], obj_c_int, radius) shapes_[k] = norm_shape.reshape((1, -1)) center_offsets[k] = mass_center - obj_c codes_[k], _ = fast_ista(norm_shape.reshape((1, -1)), self.dictionary, lmbda=self.sparse_alpha, max_iter=60) contour_std_[k] = np.sqrt(np.sum(contour_std**2)) w_h_[k] = 1. * w, 1. * h # w_h_[k] = mass_center[1] - bbox[1], bbox[3] - mass_center[1], \ # mass_center[0] - bbox[0], bbox[2] - mass_center[0] # [top, bottom, left, right] distance regs[k] = obj_c - obj_c_int # discretization error inds[k] = obj_c_int[1] * self.fmap_size['w'] + obj_c_int[0] ind_masks[k] = 1 return { 'image': img, 'shapes': shapes_, 'codes': codes_, 'offsets': center_offsets, 'std': contour_std_, 'hmap': hmap, 'w_h_': w_h_, 'regs': regs, 'inds': inds, 'ind_masks': ind_masks, 'c': center, 's': scale, 'img_id': img_id }