apdepth/util/image_util.py

# Copyright 2023 Bingxin Ke, ETH Zurich. All rights reserved.
# Last modified: 2024-05-24
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# --------------------------------------------------------------------------
# If you find this code useful, we kindly ask you to cite our paper in your work.
# Please find bibtex at: https://github.com/prs-eth/Marigold#-citation
# More information about the method can be found at https://marigoldmonodepth.github.io
# --------------------------------------------------------------------------


import matplotlib
import numpy as np
import torch
from torchvision.transforms import InterpolationMode
from torchvision.transforms.functional import resize


def colorize_depth_maps(
    depth_map, min_depth, max_depth, cmap="Spectral", valid_mask=None
):
    """
    Colorize depth maps.
    """
    assert len(depth_map.shape) >= 2, "Invalid dimension"

    if isinstance(depth_map, torch.Tensor):
        depth = depth_map.detach().squeeze().numpy()
    elif isinstance(depth_map, np.ndarray):
        depth = depth_map.copy().squeeze()
    # reshape to [ (B,) H, W ]
    if depth.ndim < 3:
        depth = depth[np.newaxis, :, :]

    # colorize
    cm = matplotlib.colormaps[cmap]
    depth = ((depth - min_depth) / (max_depth - min_depth)).clip(0, 1)
    img_colored_np = cm(depth, bytes=False)[:, :, :, 0:3]  # value from 0 to 1
    img_colored_np = np.rollaxis(img_colored_np, 3, 1)

    if valid_mask is not None:
        if isinstance(depth_map, torch.Tensor):
            valid_mask = valid_mask.detach().numpy()
        valid_mask = valid_mask.squeeze()  # [H, W] or [B, H, W]
        if valid_mask.ndim < 3:
            valid_mask = valid_mask[np.newaxis, np.newaxis, :, :]
        else:
            valid_mask = valid_mask[:, np.newaxis, :, :]
        valid_mask = np.repeat(valid_mask, 3, axis=1)
        img_colored_np[~valid_mask] = 0

    if isinstance(depth_map, torch.Tensor):
        img_colored = torch.from_numpy(img_colored_np).float()
    elif isinstance(depth_map, np.ndarray):
        img_colored = img_colored_np

    return img_colored


def chw2hwc(chw):
    assert 3 == len(chw.shape)
    if isinstance(chw, torch.Tensor):
        hwc = torch.permute(chw, (1, 2, 0))
    elif isinstance(chw, np.ndarray):
        hwc = np.moveaxis(chw, 0, -1)
    return hwc


def resize_max_res(
    img: torch.Tensor,
    max_edge_resolution: int,
    resample_method: InterpolationMode = InterpolationMode.BILINEAR,
) -> torch.Tensor:
    """
    Resize image to limit maximum edge length while keeping aspect ratio.

    Args:
        img (`torch.Tensor`):
            Image tensor to be resized. Expected shape: [B, C, H, W]
        max_edge_resolution (`int`):
            Maximum edge length (pixel).
        resample_method (`PIL.Image.Resampling`):
            Resampling method used to resize images.

    Returns:
        `torch.Tensor`: Resized image.
    """
    assert 4 == img.dim(), f"Invalid input shape {img.shape}"

    original_height, original_width = img.shape[-2:]
    downscale_factor = min(
        max_edge_resolution / original_width, max_edge_resolution / original_height
    )

    new_width = int(original_width * downscale_factor)
    new_height = int(original_height * downscale_factor)

    resized_img = resize(img, (new_height, new_width), resample_method, antialias=True)
    return resized_img


def get_tv_resample_method(method_str: str) -> InterpolationMode:
    resample_method_dict = {
        "bilinear": InterpolationMode.BILINEAR,
        "bicubic": InterpolationMode.BICUBIC,
        "nearest": InterpolationMode.NEAREST_EXACT,
        "nearest-exact": InterpolationMode.NEAREST_EXACT,
    }
    resample_method = resample_method_dict.get(method_str, None)
    if resample_method is None:
        raise ValueError(f"Unknown resampling method: {resample_method}")
    else:
        return resample_method