class PlaneNetDetector(): def __init__(self, options, config, checkpoint_dir=''): self.options = options self.config = config sys.path.append('../../existing_methods/') from PlaneNet.planenet_inference import PlaneNetDetector self.detector = PlaneNetDetector(predictNYU=False) return def detect(self, sample): detection_pair = [] for indexOffset in [0, ]: images, image_metas, rpn_match, rpn_bbox, gt_class_ids, gt_boxes, gt_masks, gt_parameters, gt_depth, extrinsics, planes, gt_segmentation = sample[indexOffset + 0].cuda(), sample[indexOffset + 1].numpy(), sample[indexOffset + 2].cuda(), sample[indexOffset + 3].cuda(), sample[indexOffset + 4].cuda(), sample[indexOffset + 5].cuda(), sample[indexOffset + 6].cuda(), sample[indexOffset + 7].cuda(), sample[indexOffset + 8].cuda(), sample[indexOffset + 9].cuda(), sample[indexOffset + 10].cuda(), sample[indexOffset + 11].cuda() image = (images[0].detach().cpu().numpy().transpose((1, 2, 0)) + self.config.MEAN_PIXEL)[80:560] pred_dict = self.detector.detect(image) segmentation = pred_dict['segmentation'] segmentation = np.concatenate([np.full((80, 640), fill_value=-1, dtype=np.int32), segmentation, np.full((80, 640), fill_value=-1, dtype=np.int32)], axis=0) planes = pred_dict['plane'] masks = (segmentation == np.arange(len(planes), dtype=np.int32).reshape((-1, 1, 1))).astype(np.float32) depth = pred_dict['depth'] depth = np.concatenate([np.zeros((80, 640), dtype=np.int32), depth, np.zeros((80, 640), dtype=np.int32)], axis=0) detections = np.concatenate([np.ones((len(planes), 4)), np.ones((len(planes), 2)), planes], axis=-1) detections = torch.from_numpy(detections).float().cuda() depth = torch.from_numpy(depth).unsqueeze(0).float().cuda() masks = torch.from_numpy(masks).float().cuda() detection_pair.append({'depth': depth, 'mask': masks.sum(0, keepdim=True), 'masks': masks, 'detection': detections}) continue return detection_pair
class TraditionalDetector(): def __init__(self, options, config, modelType=''): self.options = options self.config = config self.modelType = modelType if 'pred' in modelType: sys.path.append('../../') from PlaneNet.planenet_inference import PlaneNetDetector self.detector = PlaneNetDetector(predictSemantics=True) pass return def detect(self, sample): detection_pair = [] for indexOffset in [0, ]: images, image_metas, rpn_match, rpn_bbox, gt_class_ids, gt_boxes, gt_masks, gt_parameters, gt_depth, extrinsics, planes, gt_segmentation, gt_semantics = sample[indexOffset + 0].cuda(), sample[indexOffset + 1].numpy(), sample[indexOffset + 2].cuda(), sample[indexOffset + 3].cuda(), sample[indexOffset + 4].cuda(), sample[indexOffset + 5].cuda(), sample[indexOffset + 6].cuda(), sample[indexOffset + 7].cuda(), sample[indexOffset + 8].cuda(), sample[indexOffset + 9].cuda(), sample[indexOffset + 10].cuda(), sample[indexOffset + 11].cuda(), sample[indexOffset + 12].cuda() image = (images[0].detach().cpu().numpy().transpose((1, 2, 0)) + self.config.MEAN_PIXEL)[80:560] input_dict = {'image': cv2.resize(image, (256, 192))} if 'gt' in self.modelType: input_dict['depth'] = cv2.resize(gt_depth[0].detach().cpu().numpy()[80:560], (256, 192)) semantics = gt_semantics[0].detach().cpu().numpy()[80:560] input_dict['semantics'] = cv2.resize(semantics, (256, 192), interpolation=cv2.INTER_NEAREST) else: pred_dict = self.detector.detect(image) input_dict['depth'] = pred_dict['non_plane_depth'].squeeze() input_dict['semantics'] = pred_dict['semantics'].squeeze().argmax(-1) pass camera = sample[30][0].numpy() input_dict['info'] = np.array([camera[0], 0, camera[2], 0, 0, camera[1], camera[3], 0, 0, 0, 1, 0, 0, 0, 0, 1, camera[4], camera[5], 1000, 0]) np.save('test/input_dict.npy', input_dict) os.system('rm test/output_dict.npy') os.system('python plane_utils.py ' + self.modelType) output_dict = np.load('test/output_dict.npy', encoding='latin1')[()] segmentation = cv2.resize(output_dict['segmentation'], (640, 480), interpolation=cv2.INTER_NEAREST) segmentation = np.concatenate([np.full((80, 640), fill_value=-1, dtype=np.int32), segmentation, np.full((80, 640), fill_value=-1, dtype=np.int32)], axis=0) planes = output_dict['plane'] masks = (segmentation == np.arange(len(planes), dtype=np.int32).reshape((-1, 1, 1))).astype(np.float32) plane_depths = calcPlaneDepths(planes, 256, 192, camera, max_depth=10) depth = (plane_depths * (np.expand_dims(output_dict['segmentation'], -1) == np.arange(len(planes)))).sum(-1) depth = cv2.resize(depth, (640, 480), interpolation=cv2.INTER_LINEAR) depth = np.concatenate([np.zeros((80, 640)), depth, np.zeros((80, 640))], axis=0) detections = np.concatenate([np.ones((len(planes), 4)), np.ones((len(planes), 2)), planes], axis=-1) detections = torch.from_numpy(detections).float().cuda() depth = torch.from_numpy(depth).unsqueeze(0).float().cuda() masks = torch.from_numpy(masks).float().cuda() detection_pair.append({'depth': depth, 'mask': masks.sum(0, keepdim=True), 'masks': masks, 'detection': detections}) continue return detection_pair