Function get_crop_info

detrsmpl/models/utils/SMPLX.py:137–183

Get the transformation of points on the cropped image to the points on the original image.

(points,
                  img_metas,
                  scale_factor: float = 1.0,
                  crop_size: int = 256)

def get_crop_info(points,
                  img_metas,
                  scale_factor: float = 1.0,
                  crop_size: int = 256):
    """Get the transformation of points on the cropped image to the points on
    the original image."""
    device = points.device
    dtype = points.dtype
    batch_size = points.shape[0]
    # Get the image to crop transformations and bounding box sizes
    crop_transforms = []
    img_bbox_sizes = []
    for img_meta in img_metas:
        crop_transforms.append(img_meta['crop_transform'])
        img_bbox_sizes.append(img_meta['scale'].max())

    img_bbox_sizes = torch.tensor(img_bbox_sizes, dtype=dtype, device=device)

    crop_transforms = torch.tensor(crop_transforms, dtype=dtype, device=device)

    crop_transforms = torch.cat([
        crop_transforms,
        torch.tensor([0.0, 0.0, 1.0], dtype=dtype, device=device).expand(
            [batch_size, 1, 3])
    ],
                                dim=1)

    inv_crop_transforms = torch.inverse(crop_transforms)

    # center on the cropped body image
    center_body_crop, bbox_size = points_to_bbox(
        points, bbox_scale_factor=scale_factor)

    orig_bbox_size = bbox_size / crop_size * img_bbox_sizes

    # Compute the center of the crop in the original image
    center = (torch.einsum(
        'bij,bj->bi', [inv_crop_transforms[:, :2, :2], center_body_crop]) +
              inv_crop_transforms[:, :2, 2])

    return {
        'center': center.reshape(-1, 2),
        'orig_bbox_size': orig_bbox_size,
        # 'bbox_size': bbox_size.reshape(-1),
        'inv_crop_transforms': inv_crop_transforms,
        # 'center_body_crop': 2 * center_body_crop / (crop_size-1) - 1,
    }


def concat_images(images: List[torch.Tensor]):