SP-SLAM3 v1.0.0
ORB-SLAM3 with learned deep learning components replacing the handcrafted feature pipeline.
What's included
- SuperPoint (CNN-based keypoint detector + 256-dim float descriptor)
- LightGlue (attention-based learned matcher, optional)
- NetVLAD / CosPlace (learned place recognition, optional)
- Optical flow fallback for tracking recovery
- FP16 inference + GPU/CPU fallback
- DBoW3 vocabulary for SuperPoint descriptors
Benchmark (EuRoC, monocular, median of 5 runs)
| Sequence |
SP-SLAM3 |
ORB-SLAM3 |
Result |
| MH_01_easy |
0.0218 m |
0.034 m |
1.56x better |
| MH_02_easy |
0.0279 m |
0.036 m |
1.29x better |
Pre-trained Models (Release Assets)
| File |
Description |
superpoint.pt |
SuperPoint TorchScript model (5 MB) |
lightglue.pt |
LightGlue TorchScript model (46 MB) |
superpoint_voc.dbow3 |
DBoW3 vocabulary (113 MB) |
superpoint_voc.yml |
YAML vocabulary (298 MB) |
