metadata
license: cc-by-nc-4.0
tags:
- visual-odometry
- multi-camera
- pose-estimation
- streaming
StreamRig
Checkpoints and evaluation results for StreamRig: Exploiting Intra-Rig Geometry for Streaming Multi-Camera Odometry.
Files
| File | Description |
|---|---|
NCLT-StreamRig.pth |
StreamRig trained on NCLT (five cameras); config configs/nclt_streamrig.yaml |
KITTI360-StreamRig.pth |
StreamRig trained on KITTI-360 (four cameras); config configs/kitti360_streamrig.yaml |
SHA256SUMS |
Checksums of the two checkpoints |
eval_results/ |
Saved trajectories and metrics of the two models |
The checkpoints contain the trainable modules only (Rig-Resampler, Causal Bridge, Stream Pose Head). The frozen MapAnything backbone is obtained separately, as described in the code repository.
Results
| Dataset | Evaluation split | t_rel (%) | r_rel (°/100 m) | ATE-SE3 (m) |
|---|---|---|---|---|
| NCLT | 2012-02-19, 2012-08-20 | 2.77 | 1.39 | 28.4 |
| KITTI-360 | drives 0009, 0010 | 2.59 | 0.98 | 63.7 |
Both use stride 3 and 100–800 m relative-error segments.
Usage
Place the checkpoints in release_weights/ of the code repository and run
scripts/eval_nclt.sh or scripts/eval_kitti360.sh. Verify the files with
sha256sum -c SHA256SUMS
License
CC BY-NC 4.0. NCLT and KITTI-360 are subject to the terms of their dataset providers.