Exemplo n.º 1
0
    def __getitem__(self, idx):
        image_name = self._info.iloc[idx, -1]

        if len(self._images) == 0:
            image = imread(os.path.join(self._images_dir, image_name))
        else:
            image = self._images[image_name]

        image = self._ensure_standard_shape(image)
        center = image.shape[1] // 2
        mask = image[:, center:]
        image = image[:, :center]

        mask = cv2.cvtColor(mask, cv2.COLOR_BGR2GRAY)
        mask = cv2.threshold(mask, 127, 255, cv2.THRESH_BINARY)[1]
        mask = get_skeletion(mask)
        mask = cv2.dilate(mask, np.ones((5, 5)))

        mask[mask == 255] = 1

        if self._stage == "train":
            try:
                image, mask = self._aug(image, mask)
                mask *= 255
            except Exception as e:
                print("Exception", e)

        image, mask = self._random_crop(image, mask)
        image = self._transform(image)
        mask = torch.LongTensor(mask)
        mask = torch.clamp(mask, 0, 1)
        return image, mask
Exemplo n.º 2
0
   def test(self, source):
      """Test the model

      Args:
         source: Input to the model, either single image or directory containing images

      Returns:
         The generated image conditioned on the input image
      """
      split_len = 600 if self.opts.dataset == 'maps' else 256

      img = utils.normalize_images(utils.imread(source))
      img_A = img[:, :split_len, :]
      img_B = img[:, split_len:, :]
      img_A = np.expand_dims(img_A, 0)
      img_B = np.expand_dims(img_B, 0)
      if self.opts.direction == 'b2a':
        input_images = img_B
        target_images = img_A
      else:
        input_images = img_A
        target_images = img_B

      self.saver.restore(self.sess, self.opts.ckpt)
      utils.imwrite(os.path.join(
              self.opts.target_dir, 'target_image'),
              target_images[0], inv_normalize=True)
      utils.imwrite(os.path.join(
              self.opts.target_dir, 'conditional_image'),
              input_images[0], inv_normalize=True)

      print ' - Sampling generator images for different random initial noise'
      for idx in xrange(self.opts.sample_num):
        print 'Sampling #', idx
        if self.opts.noise_type == "gauss":
          code = gaussian_noise([1, self.opts.code_len])
        else:
          code = uniform_noise([1, self.opts.code_len])

        feed_dict = {
          self.is_training: False,
          self.images_A: img_A,
          self.images_B: img_B,
          self.code: code
        }
        if self.opts.model == 'bicycle':
          images = self.G_cvae.eval(session=self.sess, feed_dict=feed_dict)
          utils.imwrite(os.path.join(
                  self.opts.target_dir, 'test_cvae{}'.format(idx)),
                  images, inv_normalize=True)
          images = self.G_clr.eval(session=self.sess, feed_dict=feed_dict)
          utils.imwrite(os.path.join(
                  self.opts.target_dir, 'test_clr{}'.format(idx)),
                  images, inv_normalize=True)
        else:
          raise ValueError("Testing only possible for bicycleGAN")
Exemplo n.º 3
0
    def infer(self, image, save_path, bright_diff=0, is_grayscale=True):
        # read image
        if isinstance(image, str):
            img = imread(image, is_grayscale=is_grayscale)
        else:
            img = image
        img = cv2.resize(img, dsize=(256, 256))
        img = np.reshape(img, newshape=(img.shape[0], img.shape[1], 1))
        gen_avatar = self.sess.run(self.testB, feed_dict={self.test_A: [img]})
        if save_path is not None:
            save_images(gen_avatar + bright_diff,
                        size=[1, 1],
                        image_path=save_path)
        gen_avatar = np.reshape(gen_avatar,
                                newshape=list(gen_avatar.shape[1:-1]))
        gen_avatar = inverse_transform(gen_avatar)

        return gen_avatar
Exemplo n.º 4
0
    def __init__(self, stage, configs=None):
        root_dir = configs["data_path"]
        self._root_dir = root_dir
        self._images_dir = os.path.join(root_dir, "images")

        self._stage = stage

        self._info_path = os.path.join(root_dir, "{}.txt".format(stage))
        self._info = pd.read_csv(self._info_path,
                                 header=None,
                                 error_bad_lines=False)

        self._is_cached = configs["cached_npy"] == 1

        self._transform = transforms.Compose([
            transforms.ToPILImage(),
            transforms.ToTensor(),
        ])

        self._images = {}
        if self._is_cached and os.path.exists(
                os.path.join(root_dir, "{}.npy".format(stage))):
            self._images = np.load(os.path.join(root_dir,
                                                "{}.npy".format(stage)),
                                   allow_pickle=True).tolist()

        if len(self._images) == 0 and self._stage != "little":
            for idx in range(len(self._info)):
                image_name = self._info.iloc[idx, 0]
                if image_name in self._images:
                    continue

                print("Read {}".format(image_name))
                self._images[image_name] = imread(
                    os.path.join(self._images_dir, image_name))

            if self._is_cached:
                np.save(
                    os.path.join(self._root_dir, "{}.npy".format(self._stage)),
                    self._images,
                )
Exemplo n.º 5
0
    def test(self):
        """
        test model on the Test-Set
        """

        self.model.eval()
        self.model.requires_grad(False)

        val_f1s = {'f1_iou': [], 'f1_center': [], 'f1_width': [], 'f1_height': []}
        val_losses = {'all': [], 'center': [], 'width': [], 'height': []}

        t = time()
        for step, sample in enumerate(self.val_loader):
            hmap_true, y_true, file_name, aug_info = sample
            hmap_true = hmap_true.to(self.cnf.device)
            y_true = json.loads(y_true[0])

            hmap_pred = self.model.forward(hmap_true)

            x_true_center, x_true_width, x_true_height = hmap_true[0, 0], hmap_true[0, 1], hmap_true[0, 2]
            x_pred_center, x_pred_width, x_pred_height = hmap_pred[0, 0], hmap_pred[0, 1], hmap_pred[0, 2]

            # log center, width, height losses
            mask = torch.tensor(torch.where(x_true_height != 0, 1, 0), dtype=torch.float32)
            loss_center = self.cnf.masked_loss_c * nn.MSELoss()(x_pred_center, x_true_center)
            loss_width = self.cnf.masked_loss_w * MaskedMSELoss()(x_pred_width, x_true_width, mask=mask)
            loss_height = self.cnf.masked_loss_h * MaskedMSELoss()(x_pred_height, x_true_height, mask=mask)
            loss = loss_center + loss_width + loss_height
            val_losses['all'].append(loss.item())
            val_losses['center'].append(loss_center.item())
            val_losses['width'].append(loss_width.item())
            val_losses['height'].append(loss_height.item())

            y_center = [(coord[0], coord[1], coord[2]) for coord in y_true]
            y_width = [(coord[0], coord[1], coord[2], coord[3]) for coord in y_true]
            y_height = [(coord[0], coord[1], coord[2], coord[4]) for coord in y_true]

            y_center_pred = utils.local_maxima_3d(heatmap=x_pred_center, threshold=0.1, device=self.cnf.device)
            y_width_pred = []
            y_height_pred = []
            bboxes_info_pred = []
            for center_coord in y_center_pred:  # y_center_pred
                cam_dist, y2d, x2d = center_coord

                width = float(x_pred_width[cam_dist, y2d, x2d])
                height = float(x_pred_height[cam_dist, y2d, x2d])

                # denormalize width and height
                width = int(round(width * STD_DEV_WIDTH + MEAN_WIDTH))
                height = int(round(height * STD_DEV_HEIGHT + MEAN_HEIGHT))
                # width = int(round(width * MAX_WIDTH))
                # height = int(round(height * MAX_HEIGHT))

                y_width_pred.append((*center_coord, width))
                y_height_pred.append((*center_coord, height))

                x2d, y2d, cam_dist = utils.rescale_to_real(x2d=x2d, y2d=y2d, cam_dist=cam_dist, q=self.cnf.q)
                bboxes_info_pred.append((x2d - width / 2, y2d - height / 2, width, height, cam_dist))

            y_center_true = utils.local_maxima_3d(heatmap=x_true_center, threshold=0.1, device=self.cnf.device)
            bboxes_info_true = []
            for center_coord in y_center_true:

                cam_dist, y2d, x2d = center_coord

                width = float(x_true_width[cam_dist, y2d, x2d])
                height = float(x_true_height[cam_dist, y2d, x2d])

                # denormalize width and height
                width = int(round(width * STD_DEV_WIDTH + MEAN_WIDTH))
                height = int(round(height * STD_DEV_HEIGHT + MEAN_HEIGHT))
                # width = int(round(width * MAX_WIDTH))
                # height = int(round(height * MAX_HEIGHT))

                y_width_pred.append((*center_coord, width))
                y_height_pred.append((*center_coord, height))

                x2d, y2d, cam_dist = utils.rescale_to_real(x2d=x2d, y2d=y2d, cam_dist=cam_dist, q=self.cnf.q)
                bboxes_info_true.append((x2d - width / 2, y2d - height / 2, width, height, cam_dist))

            metrics_iou = compute_det_metrics_iou(bboxes_a=bboxes_info_pred, bboxes_b=bboxes_info_true)
            metrics_center = joint_det_metrics(points_pred=y_center_pred, points_true=y_center, th=1)
            metrics_width = joint_det_metrics(points_pred=y_width_pred, points_true=y_width, th=1)
            metrics_height = joint_det_metrics(points_pred=y_height_pred, points_true=y_height, th=1)
            f1_iou = metrics_iou['f1']
            f1_center = metrics_center['f1']
            f1_width = metrics_width['f1']
            f1_height = metrics_height['f1']
            val_f1s['f1_iou'].append(f1_iou)
            val_f1s['f1_center'].append(f1_center)
            val_f1s['f1_width'].append(f1_width)
            val_f1s['f1_height'].append(f1_height)

            if step < 3:
                img_original = np.array(utils.imread(self.cnf.mot_synth_path / file_name[0]).convert("RGB"))
                hmap_pred = hmap_pred.squeeze()
                out_path = self.cnf.exp_log_path / f'{step}_center_pred.mp4'
                utils.save_3d_hmap(hmap=hmap_pred[0, ...], path=out_path)
                out_path = self.cnf.exp_log_path / f'{step}_width_pred.mp4'
                utils.save_3d_hmap(hmap=hmap_pred[1, ...], path=out_path, shift_values=True)
                out_path = self.cnf.exp_log_path / f'{step}_height_pred.mp4'
                utils.save_3d_hmap(hmap=hmap_pred[2, ...], path=out_path, shift_values=True)
                out_path = self.cnf.exp_log_path / f'{step}_bboxes_pred.jpg'
                utils.save_bboxes(img_original, bboxes_info_pred, path=out_path, use_z=True, half_images=True)

                hmap_true = hmap_true.squeeze()
                out_path = self.cnf.exp_log_path / f'{step}_center_true.mp4'
                utils.save_3d_hmap(hmap=hmap_true[0, ...], path=out_path)
                out_path = self.cnf.exp_log_path / f'{step}_width_true.mp4'
                utils.save_3d_hmap(hmap=hmap_true[1, ...], path=out_path, shift_values=True)
                out_path = self.cnf.exp_log_path / f'{step}_height_true.mp4'
                utils.save_3d_hmap(hmap=hmap_true[2, ...], path=out_path, shift_values=True)
                out_path = self.cnf.exp_log_path / f'{step}_bboxes_true.jpg'
                utils.save_bboxes(img_original, bboxes_info_true, path=out_path, use_z=True, half_images=True)

            if step >= self.cnf.test_len - 1:
                break

        # log average f1 on test set
        mean_val_loss = np.mean(val_losses['all'])
        mean_val_f1_iou = np.mean(val_f1s['f1_iou'])
        mean_val_f1_center = np.mean(val_f1s['f1_center'])
        mean_val_f1_width = np.mean(val_f1s['f1_width'])
        mean_val_f1_height = np.mean(val_f1s['f1_height'])
        mean_val_loss_center = np.mean(val_losses['center'])
        mean_val_loss_width = np.mean(val_losses['width'])
        mean_val_loss_height = np.mean(val_losses['height'])
        print(f'[TEST] AVG-Loss: {mean_val_loss:.6f}, '
              f'AVG-F1_iou: {mean_val_f1_iou:.6f}, '
              f'AVG-F1_center: {mean_val_f1_center:.6f}, '
              f'AVG-F1_width: {mean_val_f1_width:.6f}, '
              f'AVG-F1_height: {mean_val_f1_height:.6f}'
              f' │ Test time: {time() - t:.2f} s')
        self.sw.add_scalar(tag='val_F1/iou', scalar_value=mean_val_f1_iou, global_step=self.current_epoch)
        self.sw.add_scalar(tag='val_F1/center', scalar_value=mean_val_f1_center, global_step=self.current_epoch)
        self.sw.add_scalar(tag='val_F1/width', scalar_value=mean_val_f1_width, global_step=self.current_epoch)
        self.sw.add_scalar(tag='val_F1/height', scalar_value=mean_val_f1_height, global_step=self.current_epoch)
        self.sw.add_scalar(tag='val_loss', scalar_value=mean_val_loss, global_step=self.current_epoch)
        self.sw.add_scalar(tag='val_loss/center', scalar_value=mean_val_loss_center, global_step=self.current_epoch)
        self.sw.add_scalar(tag='val_loss/width', scalar_value=mean_val_loss_width, global_step=self.current_epoch)
        self.sw.add_scalar(tag='val_loss/height', scalar_value=mean_val_loss_height, global_step=self.current_epoch)

        # save best model
        if self.best_val_f1 is None or mean_val_f1_iou < self.best_val_f1:
            self.best_val_f1 = mean_val_f1_iou
            torch.save(self.model.state_dict(), self.log_path / 'best.pth')
Exemplo n.º 6
0
    args = parser.parse_args()
    if not args.image_a_path:
        parser.error('-image_a_path is required.')
    if not args.image_b_path:
        parser.error('-image_b_path is required.')
    if not args.ref_lines_path:
        parser.error('-ref_lines_path is required.')
    if not args.N:
        parser.error('-N is required')

    image_a_path = args.image_a_path
    image_b_path = args.image_b_path
    ref_lines_path = args.ref_lines_path

    # Reading the images
    image_a = imread(filename=image_a_path)
    image_b = imread(filename=image_b_path)
    if image_a.shape[2] > 3:
        image_a = image_a[:, :, 0:3]
    if image_b.shape[2] > 3:
        image_a = image_a[:, :, 0:3]

    shape_a = image_a.shape
    shape_b = image_b.shape

    # n_images = args.N
    n_images = 11
    buf_shape = []
    for i in range(3):
        buf_shape.append(max(shape_a[i], shape_b[i]))
Exemplo n.º 7
0
def main():
    MAX_WIDTH = 1919
    MAX_HEIGHT = 1079
    from test_metrics import joint_det_metrics, compute_det_metrics_iou
    import json
    import numpy as np
    from torch.utils.data import DataLoader
    from conf import Conf
    from dataset.mot_synth_det_ds import MOTSynthDetDS
    from utils import utils
    import torch

    cnf = Conf(exp_name='vha_d_debug', preload_checkpoint=False)

    # load dataset
    mode = 'test'
    ds = MOTSynthDetDS(mode=mode, cnf=cnf)
    loader = DataLoader(dataset=ds, batch_size=1, num_workers=0, shuffle=False)

    # load model
    from models.vha_det_variable_versions import Autoencoder as AutoencoderVariableVersions
    model = AutoencoderVariableVersions(vha_version=1).to(cnf.device)
    model.eval()
    model.requires_grad(False)
    if cnf.model_weights is not None:
        model.load_state_dict(torch.load(cnf.exp_log_path / 'best.pth',
                                         map_location=torch.device('cpu')),
                              strict=False)

    # ======== MAIN LOOP ========
    for i, sample in enumerate(loader):
        x, y, file_name, aug_info = None, None, None, None

        if mode == 'test':
            x, y, file_name, aug_info = sample
            y_true = json.loads(y[0])
        if mode == 'train':
            x, file_name, aug_info = sample
        x = x.to(cnf.device)
        x_center, x_width, x_height = x[0, 0], x[0, 1], x[0, 2]

        y_pred = model.forward(x)
        x_pred_center, x_pred_width, x_pred_height = y_pred[0, 0], y_pred[
            0, 1], y_pred[0, 2]

        if mode == 'test':
            y = json.loads(y[0])
            y_center = [(coord[0], coord[1], coord[2]) for coord in y]
            y_width = [(coord[0], coord[1], coord[2], coord[3]) for coord in y]
            y_height = [(coord[0], coord[1], coord[2], coord[4])
                        for coord in y]

        # utils.visualize_3d_hmap(x[0, 2])
        y_center_pred = utils.local_maxima_3d(heatmap=x_pred_center,
                                              threshold=0.1,
                                              device=cnf.device)
        y_width_pred = []
        y_height_pred = []
        bboxes_info_pred = []
        # w_min = min([float(x_width[cam_dist, y2d, x2d]) for cam_dist, y2d, x2d in y_center_pred])
        # w_max = max([float(x_width[cam_dist, y2d, x2d]) for cam_dist, y2d, x2d in y_center_pred])
        # h_min = min([float(x_height[cam_dist, y2d, x2d]) for cam_dist, y2d, x2d in y_center_pred])
        # h_max = max([float(x_height[cam_dist, y2d, x2d]) for cam_dist, y2d, x2d in y_center_pred])
        for cam_dist, y2d, x2d in y_center:
            width = float(x_pred_width[cam_dist, y2d, x2d])
            height = float(x_pred_height[cam_dist, y2d, x2d])

            # denormalize width and height
            width = int(round(width * MAX_WIDTH))
            height = int(round(height * MAX_HEIGHT))

            y_width_pred.append((cam_dist, y2d, x2d, width))
            y_height_pred.append((cam_dist, y2d, x2d, height))

            x2d, y2d, cam_dist = utils.rescale_to_real(x2d=x2d,
                                                       y2d=y2d,
                                                       cam_dist=cam_dist,
                                                       q=cnf.q)
            bboxes_info_pred.append(
                (x2d - width / 2, y2d - height / 2, width, height, cam_dist))

        img_original = np.array(
            utils.imread(cnf.mot_synth_path / file_name[0]).convert("RGB"))
        if mode == 'test':
            bboxes_info_true = []
            for cam_dist, y2d, x2d, width, height in y:
                x2d, y2d, cam_dist = utils.rescale_to_real(x2d=x2d,
                                                           y2d=y2d,
                                                           cam_dist=cam_dist,
                                                           q=cnf.q)
                bboxes_info_true.append((x2d - width / 2, y2d - height / 2,
                                         width, height, cam_dist))

            metrics_iou = compute_det_metrics_iou(bboxes_info_pred,
                                                  bboxes_info_true)
            metrics_center = joint_det_metrics(points_pred=y_center_pred,
                                               points_true=y_center,
                                               th=1)
            metrics_width = joint_det_metrics(points_pred=y_width_pred,
                                              points_true=y_width,
                                              th=1)
            metrics_height = joint_det_metrics(points_pred=y_height_pred,
                                               points_true=y_height,
                                               th=1)
            f1_iou = metrics_iou['f1']
            f1_center = metrics_center['f1']
            f1_width = metrics_width['f1']
            f1_height = metrics_height['f1']
            print(
                f'f1_iou={f1_iou}, f1_center={f1_center}, f1_width={f1_width}, f1_height={f1_height}'
            )

            # for cam_dist, y2d, x2d, width, height in y_true:
            #    x2d, y2d, cam_dist = utils.rescale_to_real(x2d=x2d, y2d=y2d, cam_dist=cam_dist, q=cnf.q)
            #    bboxes_info_true.append((x2d - width / 2, y2d - height / 2, width, height, cam_dist))

            # utils.visualize_bboxes(img_original, bboxes_info_true, use_z=True, half_images=False, aug_info=aug_info,
            #                       normalize_z=False)

        # print(f'({i}) Dataset example: x.shape={tuple(x.shape)}, y={y}')

        utils.visualize_bboxes(img_original,
                               bboxes_info_pred,
                               use_z=True,
                               half_images=True,
                               aug_info=aug_info,
                               normalize_z=False)
Exemplo n.º 8
0
def main():
    from test_metrics import joint_det_metrics, compute_det_metrics_iou
    cnf = Conf(exp_name='debug')

    # load dataset
    mode = 'val'
    ds = MOTSynthDetDS(mode=mode, cnf=cnf)
    loader = DataLoader(dataset=ds,
                        batch_size=1,
                        num_workers=1,
                        shuffle=False,
                        worker_init_fn=MOTSynthDetDS.wif_test)

    # load model
    # from models.vha_det_c3d_pretrained import Autoencoder as AutoencoderC3dPretrained
    # model = AutoencoderC3dPretrained(hmap_d=cnf.hmap_d, legacy_pretrained=cnf.saved_epoch == 0).to(cnf.device)
    # model.eval()
    # model.requires_grad(False)
    # if cnf.model_weights is not None:
    #     model.load_state_dict(cnf.model_weights, strict=False)

    # ======== MAIN LOOP ========
    for i, sample in enumerate(loader):
        x, y, file_name, aug_info = None, None, None, None

        if mode == 'val' or mode == 'test':
            x, y, file_name, aug_info = sample
            y_true = json.loads(y[0])
        if mode == 'train':
            x, file_name, aug_info = sample
        x = x.to(cnf.device)
        x_center, x_width, x_height = x[0, 0], x[0, 1], x[0, 2]

        # y_pred = model.forward(x)
        # x_pred_center, x_pred_width, x_pred_height = y_pred[0, 0], y_pred[0, 1], y_pred[0, 2]

        if mode == 'test':
            y = json.loads(y[0])
            y_center = [(coord[0], coord[1], coord[2]) for coord in y]
            y_width = [(coord[0], coord[1], coord[2], coord[3]) for coord in y]
            y_height = [(coord[0], coord[1], coord[2], coord[4])
                        for coord in y]

        # utils.visualize_3d_hmap(x[0, 2])
        y_center_pred = utils.local_maxima_3d(heatmap=x_center,
                                              threshold=0.1,
                                              device=cnf.device)
        y_width_pred = []
        y_height_pred = []
        bboxes_info_pred = []
        # w_min = min([float(x_width[cam_dist, y2d, x2d]) for cam_dist, y2d, x2d in y_center_pred])
        # w_max = max([float(x_width[cam_dist, y2d, x2d]) for cam_dist, y2d, x2d in y_center_pred])
        # h_min = min([float(x_height[cam_dist, y2d, x2d]) for cam_dist, y2d, x2d in y_center_pred])
        # h_max = max([float(x_height[cam_dist, y2d, x2d]) for cam_dist, y2d, x2d in y_center_pred])
        for cam_dist, y2d, x2d in y_center_pred:
            width = float(x_width[cam_dist, y2d, x2d])
            height = float(x_height[cam_dist, y2d, x2d])

            # denormalize width and height
            # width = int(round(width * MAX_WIDTH))
            # height = int(round(height * MAX_HEIGHT))
            width = int(round(width * STD_DEV_WIDTH + MEAN_WIDTH))
            height = int(round(height * STD_DEV_HEIGHT + MEAN_HEIGHT))

            y_width_pred.append((cam_dist, y2d, x2d, width))
            y_height_pred.append((cam_dist, y2d, x2d, height))

            x2d, y2d, cam_dist = utils.rescale_to_real(x2d=x2d,
                                                       y2d=y2d,
                                                       cam_dist=cam_dist,
                                                       q=cnf.q)
            bboxes_info_pred.append(
                (x2d - width / 2, y2d - height / 2, width, height, cam_dist))

        img_original = np.array(
            utils.imread(cnf.mot_synth_path / file_name[0]).convert("RGB"))
        if mode == 'test':
            bboxes_info_true = []
            for cam_dist, y2d, x2d, width, height in y:
                x2d, y2d, cam_dist = utils.rescale_to_real(x2d=x2d,
                                                           y2d=y2d,
                                                           cam_dist=cam_dist,
                                                           q=cnf.q)
                bboxes_info_true.append((x2d - width / 2, y2d - height / 2,
                                         width, height, cam_dist))

            metrics_iou = compute_det_metrics_iou(bboxes_info_pred,
                                                  bboxes_info_true)
            metrics_center = joint_det_metrics(points_pred=y_center_pred,
                                               points_true=y_center,
                                               th=1)
            metrics_width = joint_det_metrics(points_pred=y_width_pred,
                                              points_true=y_width,
                                              th=1)
            metrics_height = joint_det_metrics(points_pred=y_height_pred,
                                               points_true=y_height,
                                               th=1)
            f1_iou = metrics_iou['f1']
            f1_center = metrics_center['f1']
            f1_width = metrics_width['f1']
            f1_height = metrics_height['f1']
            print(
                f'f1_iou={f1_iou}, f1_center={f1_center}, f1_width={f1_width}, f1_height={f1_height}'
            )

            # for cam_dist, y2d, x2d, width, height in y_true:
            #    x2d, y2d, cam_dist = utils.rescale_to_real(x2d=x2d, y2d=y2d, cam_dist=cam_dist, q=cnf.q)
            #    bboxes_info_true.append((x2d - width / 2, y2d - height / 2, width, height, cam_dist))

            # utils.visualize_bboxes(img_original, bboxes_info_true, use_z=True, half_images=False, aug_info=aug_info,
            #                       normalize_z=False)

        # print(f'({i}) Dataset example: x.shape={tuple(x.shape)}, y={y}')

        utils.visualize_bboxes(img_original,
                               bboxes_info_pred,
                               use_z=True,
                               half_images=True,
                               aug_info=aug_info,
                               normalize_z=False)
Exemplo n.º 9
0
 def get_frame(self, file_path):
     # read input frame
     frame_path = self.cnf.mot_synth_path / file_path
     frame = utils.imread(frame_path).convert('RGB')
     # frame = transforms.ToTensor()(frame)
     return frame
Exemplo n.º 10
0
    parser.add_argument("-ro", type=int, default=5)

    parser.add_argument("-closing", action="store_true")
    parser.add_argument("-rc", type=int, default=5)

    parser.add_argument("-rm_ts", type=int, default=100)

    parser.add_argument("-r_error_th", type=int, default=3)
    parser.add_argument("-output_file", type=str, default='output.txt')
    args = parser.parse_args()
    if not args.image_path:
        parser.error('-image_path is required.')
    img_path = args.image_path

    # Opening the image
    image = utils.imread(filename=img_path, as_gray=False)

    # Separating layers
    red_image = image[:, :, 0]
    green_image = image[:, :, 1]
    blue_image = image[:, :, 2]
    # Yellow layer
    yellow_image = red_image.astype(np.float64) + green_image.astype(
        np.float64) - blue_image.astype(np.float64)
    yellow_image = utils.to_uint8(yellow_image / (255.0 + 255.0))

    # Denoising
    d_lambda = args.d_lambda
    if d_lambda > 0:
        den_image = (restoration.denoise_tv_chambolle(
            image=yellow_image, weight=d_lambda) * 255).astype(np.uint8)