Function prepare_input

demo.py:119–231

Prepare input views for inference from a list of image paths. Args: img_paths (list): List of image file paths. img_mask (list of bool): Flags indicating valid images. size (int): Target image size. raymaps (list, optional): List of ray maps. raymap_

(
    img_paths, img_mask, size, raymaps=None, raymap_mask=None, revisit=1, update=True, reset_interval=10000
)

def prepare_input(
    img_paths, img_mask, size, raymaps=None, raymap_mask=None, revisit=1, update=True, reset_interval=10000
):
    """
    Prepare input views for inference from a list of image paths.

    Args:
        img_paths (list): List of image file paths.
        img_mask (list of bool): Flags indicating valid images.
        size (int): Target image size.
        raymaps (list, optional): List of ray maps.
        raymap_mask (list, optional): Flags indicating valid ray maps.
        revisit (int): How many times to revisit each view.
        update (bool): Whether to update the state on revisits.

    Returns:
        list: A list of view dictionaries.
    """
    # Import image loader (delayed import needed after adding ckpt path).
    from src.dust3r.utils.image import load_images

    images = load_images(img_paths, size=size)
    views = []

    if raymaps is None and raymap_mask is None:
        # Only images are provided.
        for i in range(len(images)):
            view = {
                "img": images[i]["img"],
                "ray_map": torch.full(
                    (
                        images[i]["img"].shape[0],
                        6,
                        images[i]["img"].shape[-2],
                        images[i]["img"].shape[-1],
                    ),
                    torch.nan,
                ),
                "true_shape": torch.from_numpy(images[i]["true_shape"]),
                "idx": i,
                "instance": str(i),
                "camera_pose": torch.from_numpy(np.eye(4, dtype=np.float32)).unsqueeze(
                    0
                ),
                "img_mask": torch.tensor(True).unsqueeze(0),
                "ray_mask": torch.tensor(False).unsqueeze(0),
                "update": torch.tensor(True).unsqueeze(0),
                "reset": torch.tensor((i+1) % reset_interval == 0).unsqueeze(0),
            }
            views.append(view)
            if (i+1) % reset_interval == 0:
                overlap_view = deepcopy(view)
                overlap_view["reset"] = torch.tensor(False).unsqueeze(0)
                views.append(overlap_view)
    else:
        # Combine images and raymaps.
        num_views = len(images) + len(raymaps)
        assert len(img_mask) == len(raymap_mask) == num_views