tools.py

import cv2
import numpy as np
import torch
from einops import rearrange

from midas import MiDaSInference

model = None

def HWC3(x):
    assert x.dtype == np.uint8
    if x.ndim == 2:
        x = x[:, :, None]
    assert x.ndim == 3
    H, W, C = x.shape
    assert C == 1 or C == 3 or C == 4
    if C == 3:
        return x
    if C == 1:
        return np.concatenate([x, x, x], axis=2)
    if C == 4:
        color = x[:, :, 0:3].astype(np.float32)
        alpha = x[:, :, 3:4].astype(np.float32) / 255.0
        y = color * alpha + 255.0 * (1.0 - alpha)
        y = y.clip(0, 255).astype(np.uint8)
        return y


def apply_midas(input_image):
    global model

    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

    if model is None:
        model = MiDaSInference(model_type="dpt_large", device=device)

    assert input_image.ndim == 3
    image_depth = input_image
    with torch.no_grad():
        image_depth = torch.from_numpy(image_depth).float().to(device)
        image_depth = image_depth / 127.5 - 1.0
        image_depth = rearrange(image_depth, 'h w c -> 1 c h w')
        depth = model(image_depth)[0]

        depth_pt = depth.clone()
        depth_pt -= torch.min(depth_pt)
        depth_pt /= torch.max(depth_pt)
        depth_pt = depth_pt.cpu().numpy()
        depth_image = (depth_pt * 255.0).clip(0, 255).astype(np.uint8)
        return depth_image


def pad64(x):
    return int(np.ceil(float(x) / 64.0) * 64 - x)


def safer_memory(x):
    # Fix many MAC/AMD problems
    return np.ascontiguousarray(x.copy()).copy()


def resize_image_with_pad(input_image, resolution, skip_hwc3=False):
    if skip_hwc3:
        img = input_image
    else:
        img = HWC3(input_image)
    H_raw, W_raw, _ = img.shape
    k = float(resolution) / float(min(H_raw, W_raw))
    interpolation = cv2.INTER_CUBIC if k > 1 else cv2.INTER_AREA
    H_target = int(np.round(float(H_raw) * k))
    W_target = int(np.round(float(W_raw) * k))
    img = cv2.resize(img, (W_target, H_target), interpolation=interpolation)
    H_pad, W_pad = pad64(H_target), pad64(W_target)
    img_padded = np.pad(img, [[0, H_pad], [0, W_pad], [0, 0]], mode='edge')

    def remove_pad(x):
        return safer_memory(x[:H_target, :W_target])

    return safer_memory(img_padded), remove_pad