feat(cameracalibrate): --check for degraded camera calibration#3118
feat(cameracalibrate): --check for degraded camera calibration#3118AaryanAgrawal wants to merge 2 commits into
Conversation
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Greptile SummaryThis PR adds live camera-calibration checks and expands the calibration workflow. The main changes are:
Confidence Score: 5/5The latest updates appear safe to merge.
Important Files Changed
Reviews (2): Last reviewed commit: "style(cameracalibrate): ruff format" | Re-trigger Greptile |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| if ( | ||
| camera_info.width | ||
| and camera_info.height | ||
| and (int(camera_info.width), int(camera_info.height)) != image_size_wh | ||
| ): | ||
| result["resolution_warning"] = ( | ||
| f"deployed CameraInfo is {camera_info.width}x{camera_info.height} but frames are " | ||
| f"{w0}x{h0}; reprojection RMS is not meaningful across resolutions." | ||
| ) |
There was a problem hiding this comment.
When the frame size differs from the deployed CameraInfo, this path only records a warning and still calculates a verdict with unscaled pixel intrinsics. For example, checking 640×480 frames against the default 1280×720 Go2 calibration can return a misleading OK from invalid reprojection and drift measurements.
| board_type = BoardType(board) | ||
| fresh_model = "plumb_bob" if board_type is BoardType.charuco else result["distortion_model"] |
There was a problem hiding this comment.
A ChArUco check of an equidistant camera always writes the accepted replacement as plumb_bob, because its drift solve also uses the pinhole ChArUco calibrator. On the default Go2 fisheye path, accepting the prompt therefore installs a calibration for the wrong projection model and downstream undistortion uses incompatible K/D semantics.
| frames_out.mkdir(parents=True, exist_ok=True) | ||
| pad = max(3, len(str(max(len(frames) - 1, 0)))) | ||
| paths: list[Path] = [] | ||
| for i, frame in enumerate(frames): | ||
| p = frames_out / f"frame_{i:0{pad}d}.png" | ||
| if not cv2.imwrite(str(p), np.asarray(frame)): |
There was a problem hiding this comment.
Writing to an existing --frames-out directory overwrites matching names but leaves higher-numbered files from earlier runs. If a 20-frame capture is followed by a 10-frame capture, a later --source folder run loads the 10 new and 10 stale images together, which can mix cameras or board geometries into one calibration.
| cand = np.asarray(candidate_corners, dtype=np.float64).reshape(-1, 2) | ||
| nearest_mean_disp_px = float("inf") | ||
| for accepted in accepted_corners_list: | ||
| acc = np.asarray(accepted, dtype=np.float64).reshape(-1, 2) | ||
| if acc.shape != cand.shape: | ||
| continue # different corner count -> not comparable, cannot prove a duplicate | ||
| mean_disp_px = float(np.mean(np.linalg.norm(cand - acc, axis=1))) |
There was a problem hiding this comment.
Partial ChArUco detections can contain different corner IDs even when their arrays have the same length, but this code subtracts them by position; detections with different lengths are treated as novel without comparison. A stationary board whose visible markers change can therefore pass the default diversity gate repeatedly and produce a poorly constrained calibration.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| @@ -1003,6 +2958,8 @@ def calibrate( | |||
| f"Detected pattern: {tuple(result.get('pattern_size', (cols, rows)))} " | |||
| f"({result.get('pattern_label', 'requested inner corners')})" | |||
| ) | |||
| if "frames_dir" in result: | |||
| typer.echo(f"Saved {result.get('n_frames_saved', 0)} frame(s) to {result['frames_dir']}") | |||
There was a problem hiding this comment.
This file is very long now. It would be better to split it.
2 participants
feat(cameracalibrate): a
--checkflag for degraded camera calibrationContribution path
Problem
CameraInfoon a live camera —cameracalibratecould write a new calib but never score the shipped one.Solution
cameracalibrate --checkscores a deployedCameraInfoon the boards it sees: reprojection RMS + intrinsics drift vs a fresh calib from the same frames →OK/DEGRADED.DEGRADED, interactivey/Nwrites the fresh calib from those same frames — no second capture.--distortion-model fisheye→ true equidistant viacv2.fisheye.calibrate); frame-diversity gate;--frames-outto save accepted frames.--check.How to Test
--checkoverlcm:/color_imagevs a ChArUco board → verdict OK, RMS 0.29 px median / 0.68 p90, drift 3.9% (n=20).uv run pytest dimos/utils/cli/cameracalibrate/Evidence
Checklist
uv run pytestfor the files changed.