Image Segmentation
PyTorch
English
painting-vision-robotics-kit
semantic-segmentation
robotics
edge-ai
construction-ai
autonomous-painting
wall-painting-robot
paint-coverage-estimation
building-facade
drywall
skirting-detection
window-detection
lidar
depth-validation
segformer
deeplabv3
Eval Results (legacy)
Update code: real-machine G-code adapter, commissioning gate, machine_job runner, pump watchdog firmware, twin completion gate, realism and masking modules (card and weights unchanged)
4eecd03 verified Download drone_align.py from constructelligence/painting-vision-robotics-kit: direct link, hf CLI and curl.
- Browser
- Download file 2.97 kB
-
https://huggingface.co/constructelligence/painting-vision-robotics-kit/resolve/main/drone_align.py
- Command line
-
hf download hf://constructelligence/painting-vision-robotics-kit/drone_align.py
-
curl -L -o drone_align.py https://huggingface.co/constructelligence/painting-vision-robotics-kit/resolve/main/drone_align.py
2.97 kB
| """Line a re-scan photo up with the photo a job was planned on, so a new camera read lands on the same wall pixels. | |
| A drone or a person comes back to the wall and takes another photo: it is never from exactly the same place. Features | |
| (AKAZE) matched between the two photos, a RANSAC homography, and the wall is planar, so one homography maps the whole | |
| wall. The new photo's segmentation is warped into the job photo's frame; pixels the new photo does not cover are | |
| "not observed" (never "bare"). Refuses, with the reason, when the photos do not match well enough to trust the warp. | |
| """ | |
| from __future__ import annotations | |
| import cv2 | |
| import numpy as np | |
| MIN_MATCHES = 30 | |
| MIN_INLIER_SHARE = 0.4 | |
| def _gray(rgb, side=1400): | |
| g = cv2.cvtColor(np.asarray(rgb), cv2.COLOR_RGB2GRAY) | |
| k = min(1.0, side / max(g.shape)) | |
| return (cv2.resize(g, (round(g.shape[1] * k), round(g.shape[0] * k)), interpolation=cv2.INTER_AREA) if k < 1 else g), k | |
| def homography(job_rgb, new_rgb): | |
| """3x3 H mapping new-photo pixels -> job-photo pixels, plus quality. Raises ValueError when it cannot be trusted.""" | |
| a, ka = _gray(job_rgb) | |
| b, kb = _gray(new_rgb) | |
| ak = cv2.AKAZE_create() | |
| k0, d0 = ak.detectAndCompute(a, None) | |
| k1, d1 = ak.detectAndCompute(b, None) | |
| if d0 is None or d1 is None or len(k0) < MIN_MATCHES or len(k1) < MIN_MATCHES: | |
| raise ValueError("too little texture to line the photos up (plain wall, blur or darkness): re-shoot with the wall edges in view") | |
| pairs = cv2.BFMatcher(cv2.NORM_HAMMING).knnMatch(d1, d0, k=2) | |
| good = [m for m, n in (p for p in pairs if len(p) == 2) if m.distance < 0.8 * n.distance] | |
| if len(good) < MIN_MATCHES: | |
| raise ValueError(f"only {len(good)} matching features between the photos: is it the same wall?") | |
| src = np.float32([k1[m.queryIdx].pt for m in good]) / kb | |
| dst = np.float32([k0[m.trainIdx].pt for m in good]) / ka | |
| H, inl = cv2.findHomography(src, dst, cv2.RANSAC, 4.0) | |
| n = 0 if inl is None else int(inl.sum()) | |
| if H is None or n < MIN_MATCHES or n < MIN_INLIER_SHARE * len(good): | |
| raise ValueError(f"the photos do not agree on one wall plane ({n} of {len(good)} matches): re-shoot square-on") | |
| det = float(np.linalg.det(H[:2, :2])) | |
| if not 0.2 < det < 5.0: | |
| raise ValueError("the alignment would stretch the photo implausibly: re-shoot from closer to where the first was taken") | |
| return H, {"matches": len(good), "inliers": n, "inlier_share": round(n / len(good), 3), "scale": round(det ** 0.5, 3)} | |
| def warp_classes(classes, H, shape): | |
| """Warp a class map from the new photo into the job frame. Returns (warped classes, covered mask).""" | |
| h, w = shape | |
| warped = cv2.warpPerspective(np.asarray(classes, np.uint8), H, (w, h), flags=cv2.INTER_NEAREST, borderValue=255) | |
| covered = cv2.warpPerspective(np.full(np.asarray(classes).shape, 255, np.uint8), H, (w, h), flags=cv2.INTER_NEAREST, borderValue=0) > 0 | |
| return warped, covered | |