Function evaluate_video_error

src/evaluate.py:168–231

(
    result_images, gt_images, masks,
    flownet_checkpoint_path, evaluate_warping_error=False,
    printlog=True
)

def evaluate_video_error(
    result_images, gt_images, masks,
    flownet_checkpoint_path, evaluate_warping_error=False,
    printlog=True
):
    total_error = 0
    total_psnr = 0
    total_ssim = 0
    total_p_dist = 0
    # worker = torchvision.transforms.Resize(SIZE, Image.BILINEAR)

    for i, (result, gt, mask) in enumerate(
        zip(result_images, gt_images, masks)
    ):
        # mask = np.expand_dims(mask, 2)
        # gt = worker(gt)
        # mask = worker(mask)

        mse, ssim_value, psnr_value, p_dist = evaluate_image(gt, result)
        total_error += mse
        total_ssim += ssim_value
        total_psnr += psnr_value
        total_p_dist += p_dist
        logger.debug(
            f"Frame {i}: MSE {mse} PSNR {psnr_value} SSIM {ssim_value} "
            f"Percep. Dist. {p_dist}"
        )

    if evaluate_warping_error:
        init_warping_model(flownet_checkpoint_path)
        # These readers are lists of images
        # After np.array, they are in shape (H, W, C)
        # While the required input of the temporal_warping_error is (B, L, C, H, W)
        # So the tensors are unsqueezed and permuted into such shape
        targets = torch.Tensor(
            [np.array(x) for x in gt_images]
        ).unsqueeze(0).permute(0, 1, 4, 2, 3)
        masks = torch.Tensor(
            [np.array(x) for x in masks]
        ).unsqueeze(3).unsqueeze(0).permute(0, 1, 4, 2, 3)
        outputs = torch.Tensor([np.array(x) for x in result_images]).unsqueeze(0).permute(0, 1, 4, 2, 3)
        data_input = {
            "targets": targets,
            "masks": masks
        }
        model_output = {"outputs": outputs}
        warping_error = temporal_warping_error(data_input, model_output).cpu().item()
        if printlog:
            logger.info(f"Warping error: {warping_error}")
    else:
        warping_error = 0

    if printlog:
        logger.info(f"Avg MSE: {total_error / len(result_images)}")
        logger.info(f"Avg PSNR: {total_psnr / len(result_images)}")
        logger.info(f"Avg SSIM: {total_ssim / len(result_images)}")
        logger.info(
            f"Avg Perce. Dist.: {total_p_dist / len(result_images)}")