SemMap models
Models for SemMap, which produces per-pixel segmentation, traversability and confidence for mobile robots.
| File | What it is |
|---|---|
onnx/dinov3s_backbone.onnx |
DINOv3 ViT-S/16 feature extractor, frozen and shared |
onnx/dinov3s_sem_seg_snow_v1.* |
decoder for alpine and snow scenes (GrandTour SNOW-2) |
onnx/dinov3s_sem_seg_trail_v4.* |
decoder for forest trails |
onnx/dinov3s_jafar_*.onnx |
optional JAFAR feature sharpening |
torch/*.pth |
the trained decoders, needed only to re-export |
The ONNX files run on both x86 and Jetson. TensorRT engines are not published,
because each one only works on the GPU and TensorRT version that built it; build
them locally with run_export_all_trt.py.
Usage
git clone https://github.com/rai-inst/sem_map.git && cd sem_map
scripts/artifacts/download.sh
Licensing
The backbone and every decoder trained on it are derived from Meta's DINOv3 and
are released under the DINOv3 licence (DINOv3_LICENSE.md), not the MIT
licence covering the SemMap source code. The licence must stay with the files,
anything derived from them stays under it, and it forbids military, weapons,
nuclear, espionage and ITAR-controlled use.
The JAFAR sharpening models are Apache-2.0.