Direct RGB–LiDAR provider readiness review¶
Reviewed2026-10-30. Machine report and schema record exact repository commits, local source paths/hashes and unresolved prerequisites. Official source code was inspected in isolated /workspace/work checkouts, not copied into src/calibrex or executed. No inference, checkpoint download, accuracy claim or fresh dataset use occurred.
| Candidate | Verified official source | Mechanism | Current decision |
|---|---|---|---|
| CoFiI2P | WHU-USI3DV/CoFiI2P | Learned coarse/fine point–pixel correspondences, PnP-RANSAC | Conditional first implementation candidate; blocked prerequisites |
| CorrI2P | rsy6318/CorrI2P | Dense cross-modal correspondences | Secondary reference; not runnable as supplied here |
| DeepI2P | lijx10/DeepI2P | Frustum/fine classification with registration | Classification comparator, not selected for next pilot |
Previously rejected SuperGlue/I2PNet correspondence exports are not reopened. A guessed2D3D-MATR repository URL did not resolve; its failed lookup is retained, and no official-source or feasibility assertion is made for that method.
Source evidence and environment¶
Current runtime is torch2.4.0+cpu with no CUDA and no Open3D, about1.3GB free disk at review time. CoFiI2P README describes CUDA11.6/PyTorch1.13.1/Open3D0.17; its evaluation calls model.cuda() and moves inputs to CUDA. Its point pyramid imports Open3D torch neighbor layers. This blocks the unmodified official evaluation in the current environment, not proof that a future CPU port is impossible.
CorrI2P evaluation likewise uses CUDA and references a locally trained .t7 path. README says prepared dataset download is temporarily unavailable. The review did not establish a usable official weight artifact; this is not a claim none exists. DeepI2P has a root MIT license, but its index_max setup instantiates CUDAExtension, and its documented coarse path also uses Ceres. Its README's listed weight links are Oxford models; a suitable KITTI artifact was not established here.
No root LICENSE/COPYING is tracked in the pinned CoFiI2P/CorrI2P trees. Public availability is not assumed to authorize redistribution; terms need clarification before packaging or copying their code. This review makes no broad legal verdict. Checkpoint bytes/digests and their terms remain unverified for all candidates; links in READMEs are not reproducible checkpoint identities. No drive/archive or large prepared dataset was downloaded. CoFiI2P's102GB prepared KITTI download is not required for a future local single-batch adapter and is not attempted.
All three code paths use KITTI sequence0–8 for training and9/10 for validation in their source loaders. Exact checkpoint training provenance and mapping to our raw-frame identities remain unverified; no train/test isolation claim is made. README-reported accuracy uses different protocols, datasets and gates and is not compared numerically to Calibrex's .5deg/.20m gate.
Critical input boundary¶
CoFiI2P's data/kitti.py applies dataset calibration P_Tr to place the cloud in camera coordinates, then applies a random transform P. That is its declared benchmark construction, not evidence of misconduct. It is nevertheless incompatible with blindly invoking the loader for an unknown-extrinsic calibration comparison. Its model has separate test inference; an adapter must bypass dataset calibration and ground-truth correspondence construction entirely, providing only explicit initial T_camera_lidar, original image/intrinsics and LiDAR/intensity/normals.
Transform points/normals with the initializer, preserve point identities, map cropped/resized feature pixels back to the source camera, and compose a predicted camera-from-initial-camera correction with the initializer. Input provenance must exclude the reference artifact. Ground-truth labels, reference point filtering or dataset calibration must never enter inference preparation. Synthetic transform/ crop tests and a changed-reference invariance test are required before real input.
Concrete conditional pilot¶
Candidate priority is CoFiI2P because it exposes coarse/fine correspondences rather than the already rejected diffuse I2PNet export. This is a mechanism-based choice, not a predicted accuracy win. Before any inference, require:
- Suitable GPU runtime or a separately validated CPU port; lock all dependencies and memory/time limits. Do not silently replace CUDA neighbor operations.
- Official checkpoint bytes with SHA256, source link and applicable terms; resolve code redistribution terms if packaging is intended. Keep external code isolated.
- Camera/point preprocessing adapter with initializer-only boundary and exact coordinate/crop/normal/determinism tests; no official reference-aligned loader.
- Checkpoint training/frame provenance and scope designation. Unknown overlap permits only explicitly overlap-unknown development evidence, never holdout proof.
Then freeze one bounded spent-data pilot on the six existing drives/18 original frames and positive/negative/reference starts. Retain all54 provider invocations, per-frame support and failure denominators. Use the existing provider-neutral export/quality/PnP boundaries; do not relabel direct network pose as correspondence success. Define fit-only acceptance and spatial holdout quality before running; reference geometry scores afterward. Report source/sampled point counts, crop, geometry support, exact artifacts and all failed poses. Historical joint MI6/12 is contextual until multi-frame aggregation/input matching is frozen. No fresh holdouts or new recordings are needed for this pilot.
No candidate is executable unmodified under the present verified resource/checkpoint conditions. The next concrete work is satisfying these prerequisites, not another surface-mask variant or a runtime-as-accuracy claim. No installation or execution was performed just to hide a known dependency blocker.
Follow-up: execution/checkpoint readiness records actual proxy403 weight access, CPU-only runtime, historical training-translation correction and unresolved checkpoint/image provenance. Inference remains deferred.