Files
zed-playground/py_workspace/.sisyphus/notepads/full-icp-pipeline/learnings.md
T

55 lines
3.4 KiB
Markdown

- Corrected success gate logic to `> 0`.
- Added INFO logging for all attempted ICP pairs.
- Ensured all pairs are stored in `metrics.per_pair_results`.
- Fixed overlap skip logging to use DEBUG level.
- Fixed syntax and indentation errors in `aruco/icp_registration.py` that were causing unreachable code and malformed control flow.
- Relaxed success gate to `metrics.num_cameras_optimized > 0`, allowing single-camera optimizations to be considered successful.
- Implemented comprehensive per-pair diagnostic logging: INFO for ICP results (fitness, RMSE, convergence) and DEBUG for overlap skips.
- Ensured all attempted ICP results are stored in `metrics.per_pair_results` for better downstream diagnostics.
- Updated `tests/test_icp_registration.py` to reflect the new success gate logic.
## Task 3: 3D AABB Overlap Check
- Implemented `compute_overlap_3d` in `aruco/icp_registration.py`.
- Added `overlap_mode` to `ICPConfig` (defaulting to "xz").
- Verified 3D overlap logic with new tests in `tests/test_icp_registration.py`.
- Confirmed that empty inputs return 0.0 volume.
- Confirmed that disjoint boxes return 0.0 volume.
- Confirmed that partial and full overlaps return correct hand-calculable volumes.
## Task 2: Point Extraction & Preprocessing
- Implemented `extract_scene_points` with floor, hybrid, and full modes.
- Implemented `preprocess_point_cloud` with statistical outlier removal (SOR).
- Added `region` field to `ICPConfig` dataclass.
- Added comprehensive tests for new extraction modes and preprocessing.
- Verified backward compatibility for floor mode.
- Verified hybrid mode behavior (vertical structure inclusion and fallback).
- Verified full mode behavior.
- Verified SOR preprocessing effectiveness.
## Task 4: TukeyLoss Robust Kernel Support
- Added `robust_kernel` and `robust_kernel_k` to `ICPConfig`.
- Implemented TukeyLoss application in `pairwise_icp` for both Point-to-Plane and Generalized ICP.
- Verified that TukeyLoss correctly handles outliers in synthetic tests, maintaining convergence accuracy.
- Default behavior remains backward-compatible with `robust_kernel="none"`.
## Task 6: FPFH+RANSAC Global Pre-alignment
- Implemented `compute_fpfh_features` and `global_registration` using Open3D RANSAC.
- Added `global_init` flag to `ICPConfig` (default False).
- Integrated global registration into `refine_with_icp` as a pre-alignment step before pairwise ICP.
- Added safety checks: global registration result is only used if fitness > 0.1 and the resulting transform is within `max_rotation_deg` and `max_translation_m` bounds relative to the initial extrinsic guess.
- Verified with synthetic tests:
- `test_compute_fpfh_features`: Validates feature dimension and count.
- `test_global_registration_known_transform`: Confirms RANSAC can recover a known large transform (30 deg rotation).
- `test_refine_with_icp_global_init_success`: End-to-end test showing global init can recover from a very bad initial guess (90 deg error) where local ICP would fail.
## Task 8: Relax ICPConfig defaults
- Relaxed defaults for ICPConfig to improve convergence and allow more flexible corrections.
- New defaults:
- min_fitness: 0.15
- min_overlap_area: 0.5
- gravity_penalty_weight: 2.0
- max_correspondence_distance_factor: 2.5
- max_translation_m: 0.3
- max_rotation_deg: 10.0
- Verified with 36 passing tests and clean basedpyright (0 errors, though many warnings due to missing stubs).