depth_texture_mask.py

# depth_texture_mask.py
# Modified: lazy MiDaS init and safe for server use.

import os
import cv2
import torch
import numpy as np
import matplotlib.pyplot as plt

# Globals (initialized by init_midas)
midas = None
midas_transforms = None
transform = None
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
_midas_initialized = False

def init_midas(model_name="DPT_Hybrid", device_override=None, force_reload=False):
    """
    Initialize/load the MiDaS model and transforms into global variables.
    Call this once (e.g., at FastAPI startup).
    """
    global midas, midas_transforms, transform, device, _midas_initialized

    if device_override is not None:
        device = device_override
    else:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    if _midas_initialized and not force_reload:
        return

    # Use torch.hub to load MiDaS transforms & model
    # NOTE: this will download if not cached
    midas = torch.hub.load("intel-isl/MiDaS", model_name, pretrained=True)
    midas_transforms = torch.hub.load("intel-isl/MiDaS", "transforms")
    # choose the appropriate transform (DPT / midas small has different names)
    if hasattr(midas_transforms, "dpt_transform"):
        transform = midas_transforms.dpt_transform
    elif hasattr(midas_transforms, "small_transform"):
        transform = midas_transforms.small_transform
    else:
        # fallback: try a generic 'transform'
        transform = getattr(midas_transforms, "transform", None)

    midas.to(device).eval()
    _midas_initialized = True
    return

def _ensure_initialized():
    if not _midas_initialized:
        init_midas()

def generate_texture_depth_mask(input_data, mask_only=False):
    """
    Generate a texture + depth structural mask.

    Supports:
        - File paths (.jpg, .png)
        - NumPy arrays (H,W,C) RGB or RGBA
        - List of inputs (batch mode)

    Returns:
        mask_only=False:
            - Single: (fig, mask)
            - Batch: list of (fig, mask)

        mask_only=True:
            - Single: mask
            - Batch: list of masks
    """
    _ensure_initialized()

    def _process_single(image_source):
        # Load image (array or file path)
        if isinstance(image_source, np.ndarray):
            img_rgb = image_source
            if img_rgb.shape[-1] == 4:
                img_rgb = img_rgb[:, :, :3]
            img_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
        elif isinstance(image_source, str) and os.path.isfile(image_source):
            img_bgr = cv2.imread(image_source)
            if img_bgr is None:
                raise ValueError(f"Could not read {image_source}")
            img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
        else:
            raise TypeError("Input must be a file path or NumPy image array.")

        gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
        blurred = cv2.GaussianBlur(gray, (3, 3), 0)

        # Depth (MiDaS)
        t = transform(img_rgb).to(device)
        if t.ndim == 3:
            t = t.unsqueeze(0)

        with torch.no_grad():
            depth = midas(t)
            depth = torch.nn.functional.interpolate(
                depth.unsqueeze(1),
                size=gray.shape,
                mode="bicubic",
                align_corners=False
            ).squeeze()

        depth = depth.cpu().numpy()
        depth = cv2.normalize(depth, None, 0, 255, cv2.NORM_MINMAX)
        depth_mask = cv2.convertScaleAbs(255 - depth)

        # Texture features
        canny = cv2.Canny(blurred, 40, 120)
        lap = cv2.convertScaleAbs(cv2.Laplacian(blurred, cv2.CV_64F))

        corners = cv2.cornerHarris(np.float32(blurred), 2, 3, 0.04)
        corners = cv2.dilate(corners, None)
        corner_mask = np.zeros_like(gray)
        corner_mask[corners > 0.01 * corners.max()] = 255

        edges_all = cv2.addWeighted(canny, 0.6, lap, 0.4, 0)
        mask = cv2.bitwise_or(edges_all, corner_mask)
        mask = cv2.addWeighted(mask, 0.8, depth_mask, 0.2, 0)

        noise = np.random.randint(0, 60, gray.shape, dtype=np.uint8)
        mask = cv2.addWeighted(mask, 1.0, noise, 0.2, 0)
        mask = cv2.convertScaleAbs(mask)

        if mask_only:
            return mask

        # Visualization mode
        fig, ax = plt.subplots(1, 2, figsize=(14, 6))
        ax[0].imshow(img_rgb)
        ax[0].set_title("Original Image")
        ax[0].axis("off")

        ax[1].imshow(mask, cmap="gray")
        ax[1].set_title("Texture + Depth Structural Mask")
        ax[1].axis("off")

        plt.tight_layout()
        return fig, mask

    # Batch support
    if isinstance(input_data, list):
        return [_process_single(item) for item in input_data]

    return _process_single(input_data)

# CLI entrypoint preserved for local use
if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--input", type=str, required=True)
    parser.add_argument("--save", type=str, default="./mask_img.png")
    parser.add_argument("--mask_only", action="store_true")
    args = parser.parse_args()
    output = generate_texture_depth_mask(args.input, mask_only=args.mask_only)
    if args.mask_only:
        mask = output
    else:
        fig, mask = output
    cv2.imwrite(args.save, mask)
    
    print(f"[OK] Saved mask to {args.save}")