#!/usr/bin/env python3 """Floor detection checks. 1. Synthetic rooms (real_data test split, camera 'synthetic'): the generator draws a floor below the wall, so ground truth = non-wall pixels below the lowest wall / skirting pixel of each column. Scores the raw general model, the cleaned mask, and the wall/floor line. Synthetic floors are flat planks: this is a floor of 'is it wired up right', not a real-world accuracy figure. 2. Real photos (surface_eval / surface_eval2 interiors): no floor labels exist, so this writes a contact sheet to look at, and reports how often the floor mask overlaps what the wall model calls wall (it should be ~0). python eval_floor.py --sheet out.jpg """ from __future__ import annotations import argparse import csv import glob import json from pathlib import Path import cv2 import numpy as np import floor_detect as FD ROOT = Path(__file__).parent def synthetic_gt(mask): """Floor = the pixels below the lowest wall / skirting / object pixel of their column, that are none of those.""" h, w = mask.shape solid = np.isin(mask, (1, 2, 3, 4, 5, 6, 7, 8, 9)) has = solid.any(axis=0) lowest = np.where(has, h - 1 - np.argmax(solid[::-1], axis=0), h) return (np.arange(h)[:, None] > lowest[None, :]) & ~solid def iou(a, b): u = (a | b).sum() return float((a & b).sum() / u) if u else 1.0 def main(): ap = argparse.ArgumentParser() ap.add_argument("--limit", type=int, default=0) ap.add_argument("--sheet") ap.add_argument("--real", type=int, default=24) a = ap.parse_args() rows = [r for r in csv.DictReader(open(ROOT / "real_data" / "metadata.csv")) if r["split"] == "test" and r["camera_id"] == "synthetic"] rows.sort(key=lambda r: r["image"]) if a.limit: rows = rows[:a.limit] res = {"raw": [], "cleaned": [], "cleaned_no_walls": []} line_err = [] for r in rows: stem = r["image"].rsplit(".", 1)[0] rgb = cv2.cvtColor(cv2.imread(str(ROOT / "real_data/test/images" / r["image"])), cv2.COLOR_BGR2RGB) cls = np.asarray(cv2.imread(str(ROOT / "real_data/test/masks" / f"{stem}.png"), cv2.IMREAD_UNCHANGED)) gt = synthetic_gt(cls) prob = FD.floor_probability(rgb) res["raw"].append(iou(prob >= 0.5, gt)) res["cleaned_no_walls"].append(iou(FD.clean_floor(prob, None), gt)) m = FD.clean_floor(prob, cls) res["cleaned"].append(iou(m, gt)) fit = FD.floor_line(m)["fit"] if fit and gt.any(): ys = np.nonzero(gt.any(axis=1))[0] line_err.append(abs((fit["y_at_left"] + fit["y_at_right"]) / 2 - ys[0]) / rgb.shape[0]) out = {"synthetic_images": len(rows), **{k: round(float(np.mean(v)), 4) for k, v in res.items()}, "line_mean_abs_error_frac_of_height": round(float(np.mean(line_err)), 4) if line_err else None, "line_found_share": round(len(line_err) / max(len(rows), 1), 3)} print(json.dumps(out, indent=1)) if a.sheet: imgs = sorted(glob.glob(str(ROOT / "surface_eval2/images/*.jpg")) + glob.glob(str(ROOT / "surface_eval/images/*.jpg"))) rng = np.random.default_rng(4) pick = [imgs[i] for i in rng.permutation(len(imgs))[:a.real]] tiles, overlap = [], [] for f in pick: rgb = cv2.cvtColor(cv2.imread(f), cv2.COLOR_BGR2RGB) m, info = FD.detect_floor(rgb) vis = rgb.copy() vis[m] = (0.45 * vis[m] + 0.55 * np.array([176, 125, 79])).astype(np.uint8) for x, y in info["line"]["points"]: cv2.circle(vis, (x, y), max(2, rgb.shape[1] // 250), (255, 255, 255), -1) tiles.append(cv2.resize(vis, (320, 240))) while len(tiles) % 6: tiles.append(np.zeros((240, 320, 3), np.uint8)) grid = np.vstack([np.hstack(tiles[i:i + 6]) for i in range(0, len(tiles), 6)]) cv2.imwrite(a.sheet, cv2.cvtColor(grid, cv2.COLOR_RGB2BGR)) if __name__ == "__main__": main()