backend/app/modules/depth.py

"""Monocular depth estimation (MiDaS) with a synthetic fallback.

Real depth (when the ML stack is present) is used as inpaint/ControlNet
conditioning and to help recover the floor plane. Without ML we synthesize a
plausible vertical gradient (far near the top/horizon, near at the bottom),
which is enough for conditioning and a UI preview and is consistent with the
heuristic floor model.
"""
from __future__ import annotations

import functools
from pathlib import Path

import cv2
import numpy as np

from ..config import settings
from .ml_runtime import ml_available


@functools.lru_cache(maxsize=1)
def _load_midas():
    import torch

    model = torch.hub.load("intel-isl/MiDaS", settings.MIDAS_MODEL)
    transforms = torch.hub.load("intel-isl/MiDaS", "transforms")
    tf = (
        transforms.small_transform
        if "small" in settings.MIDAS_MODEL.lower()
        else transforms.dpt_transform
    )
    model.eval()
    return model, tf


def estimate_depth(image_path: Path) -> np.ndarray:
    """Return a HxW float32 depth map normalized to [0, 1] (1 = nearest)."""
    img = cv2.imread(str(image_path))
    if img is None:
        raise ValueError(f"Could not read image for depth: {image_path}")
    h, w = img.shape[:2]

    if not ml_available():
        return _synthetic_depth(h, w)
    try:
        import torch

        model, tf = _load_midas()
        rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        batch = tf(rgb)
        with torch.no_grad():
            pred = model(batch)
            pred = torch.nn.functional.interpolate(
                pred.unsqueeze(1), size=(h, w), mode="bicubic", align_corners=False
            ).squeeze().cpu().numpy()
        d = pred.astype("float32")
        d = (d - d.min()) / (float(np.ptp(d)) + 1e-6)
        return d
    except Exception:
        return _synthetic_depth(h, w)


def _synthetic_depth(h: int, w: int) -> np.ndarray:
    col = np.linspace(0.0, 1.0, h, dtype="float32").reshape(h, 1)
    return np.repeat(col, w, axis=1)


def depth_to_image(depth: np.ndarray) -> np.ndarray:
    """Convert a [0, 1] depth map to an 8-bit BGR image for saving/preview."""
    vis = (np.clip(depth, 0.0, 1.0) * 255).astype("uint8")
    return cv2.applyColorMap(vis, cv2.COLORMAP_INFERNO)