verify_r1_metric.py

"""R1-1 — metric depth certification (local harness; needs torch+transformers
and reference room photos — not part of `make verify`, which uses precomputed
bundles).

Runs the configured depth checkpoint on reference room photos and validates
that the output is genuinely METRIC:

  1. floor depth range plausible for an interior (p5/p95 within 0.3-20 m)
  2. ground-plane consistency: on a floor plane, inverse depth is linear in
     image row (1/Z = (y - y_horizon) / (h_cam * f)); the fit must hold
     (R^2 >= 0.9 over floor rows)
  3. absolute scale: the camera height recovered from that fit's slope
     (h = 1 / (slope * f), f ~ image width) must land in 0.7-2.5 m — the
     handheld-phone band. This is the automated equivalent of the backlog's
     "door height ~2.0 m +/-15%" check: both test absolute metric scale, but
     this one needs no manual annotation.

Usage:
    python verify_r1_metric.py <room-photo.jpg> [more photos...]
"""

import sys

import cv2
import numpy as np
import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModelForDepthEstimation

# single source of truth: read the configured model + metric predicate from app.py
src = open("app.py").read()
ns = {}
start = src.index("def depth_model_is_metric")
end = src.index("\nENABLE_DEPTH", start)
exec(compile(src[start:end], "app.py", "exec"), ns)
import re

MODEL = re.search(r'depth_model_name",\s*\n(?:\s*#.*\n)*\s*"([^"]+)"', src).group(1)
depth_model_is_metric = ns["depth_model_is_metric"]

FLOOR_FRAC = 0.45  # treat the bottom 45% of the frame as floor-dominated


def run_depth(img):
    processor = run_depth.processor
    model = run_depth.model
    inputs = processor(images=img, return_tensors="pt")
    with torch.no_grad():
        out = model(**inputs)
        depth = torch.nn.functional.interpolate(
            out.predicted_depth.unsqueeze(1),
            size=(img.height, img.width),
            mode="bicubic",
            align_corners=False,
        ).squeeze().numpy()
    return cv2.GaussianBlur(depth.astype(np.float32), (0, 0), sigmaX=3)


def main():
    photos = sys.argv[1:]
    if not photos:
        print(__doc__)
        return 2
    print(f"model: {MODEL}")
    if not depth_model_is_metric(MODEL):
        print("!! configured model is not metric — R1-1 not in effect")
        return 1

    print("loading checkpoint...")
    run_depth.processor = AutoImageProcessor.from_pretrained(MODEL)
    run_depth.model = AutoModelForDepthEstimation.from_pretrained(MODEL).eval()

    ok = True
    for path in photos:
        img = Image.open(path).convert("RGB")
        if max(img.size) > 1280:
            s = 1280 / max(img.size)
            img = img.resize((round(img.width * s), round(img.height * s)), Image.LANCZOS)
        w, h = img.size
        depth = run_depth(img)

        floor = depth[int(h * (1 - FLOOR_FRAC)):, :]
        p5, p95 = np.percentile(floor, 5), np.percentile(floor, 95)
        range_ok = 0.3 <= p5 and p95 <= 20.0

        # row-median inverse depth over the floor band; fit 1/Z = a*y + b
        ys = np.arange(int(h * (1 - FLOOR_FRAC)), h)
        inv = np.array([np.median(1.0 / np.maximum(depth[y], 0.05)) for y in ys])
        a, b = np.polyfit(ys, inv, 1)
        pred = a * ys + b
        ss_res = float(np.sum((inv - pred) ** 2))
        ss_tot = float(np.sum((inv - inv.mean()) ** 2)) + 1e-12
        r2 = 1 - ss_res / ss_tot

        focal = float(w)  # P0 convention: f ~ image width
        horizon_y = -b / a if abs(a) > 1e-12 else float("nan")
        # exact ground-plane relation for a pitched camera:
        #   1/Z = (sin(t)*f - cos(t)*y') / (h*f)   ->   h = cos(t) / (a*f)
        # with pitch t recovered from the fitted horizon row.
        pitch = np.arctan2(h / 2 - horizon_y, focal)
        cam_h = float(np.cos(pitch) / (a * focal)) if a > 1e-9 else float("inf")

        plane_ok = r2 >= 0.90 and a > 0
        height_ok = 0.7 <= cam_h <= 2.5
        passed = range_ok and plane_ok and height_ok
        ok &= passed
        print(
            f"  [{'PASS' if passed else 'FAIL'}] {path.split('/')[-1]}: "
            f"floor p5-p95 = {p5:.2f}-{p95:.2f} m | invZ-fit R2={r2:.3f} | "
            f"camera height = {cam_h:.2f} m | horizon y = {horizon_y:.0f}/{h}"
        )
        if not range_ok:
            print("        !! floor depth outside 0.3-20 m")
        if not plane_ok:
            print("        !! inverse depth not linear in row — not plane-consistent")
        if not height_ok:
            print("        !! camera height outside handheld band 0.7-2.5 m")

    print("\n" + ("ALL R1-1 METRIC CHECKS PASSED" if ok else "R1-1 METRIC CHECKS FAILED"))
    return 0 if ok else 1


if __name__ == "__main__":
    raise SystemExit(main())