NexusDwin/sailswarm-ewasr-lars

SailSwarm eWaSR (LaRS-trained water/sky/obstacle segmentation)

ewasr_resnet18 (Teršek et al.) trained on the LaRS maritime benchmark for 3-class semantic segmentation — obstacle / water / sky — for the SailSwarm autonomous-sailboat obstacle-detection module (University of Konstanz).

Used to derive a robust water edge (navigation up-vector / horizon) and the true waterline-contact point for monocular range, replacing a brittle RANSAC-Canny horizon.

Results

• Generalises to our undistorted fisheye Lake-Constance footage: mean water≈0.51 / sky≈0.37 / obstacle≈0.12, water-edge confidence ≈0.79.
• Trained 512×384; early-stopped ~epoch 26.

Intended use & limits

• Input: undistorted RGB frame. Output: per-pixel class {0 obstacle,1 water,2 sky}.
• Daytime RGB only (no thermal/night). Not yet validated for live navigation.

Provenance / license

Architecture: eWaSR (tersekmatija/eWaSR). Training data: LaRS (research/ non-commercial — derived weights inherit its terms). Verify both licenses before any commercial or redistribution use.

Citation

This model derives from the LaRS dataset/benchmark. If you use it, please cite:

@InProceedings{Zust2023LaRS,
  title={LaRS: A Diverse Panoptic Maritime Obstacle Detection Dataset and Benchmark},
  author={{\v{Z}}ust, Lojze and Per{\v{s}}, Janez and Kristan, Matej},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2023}
}