Skip to content

SiteMind v0.1.0 — Fine-tuned checkpoints

Choose a tag to compare

View all tags
@Minyue0213 Minyue0213 released this 07 Sep 11:45
· 1 commit to main since this release
v0.1.0
bb0b184

SiteMind v0.1.0 checkpoints

This release publishes the two PP-LiteSeg checkpoints trained for SiteMind's camera-based terrain perception module.

Files

best_safety.pth

  • Selected at epoch 11 by the validation safety score used in scripts/train_ppliteseg.py.
  • Recommended for the conservative camera-to-BEV risk-mapping and planning pipeline.
  • SHA-256: dbd5c83af8120136051f120ed7c114885b6c346ecdbe69ce2b05e83ea2f51430

best_miou.pth

  • Selected at epoch 13 by validation mean IoU.
  • Provided for semantic-segmentation evaluation and comparison.
  • SHA-256: 7cbee2d7019170ea549a0b78072f8cb9ab97c6732af007e2e90e150a050d2b2c

Both files are full training checkpoints containing model, optimizer, scheduler and mixed-precision states. The model uses the 64-class GOOSE ontology and a 512 × 512 input resolution.

Training configuration

  • Architecture: PP-LiteSeg B75
  • Training data: GOOSE-Ex 2D training split
  • Validation data: GOOSE-Ex 2D validation split
  • Objective: equal-weight cross-entropy and Generalized Dice loss
  • Optimizer: AdamW
  • Epochs: 15
  • Random seed: 42

The validation split was not used for parameter updates. Refer to README.md and docs/REPRODUCIBILITY.md for reported evaluation results, limitations and commands.

Attribution and terms

The checkpoints are derived from the GOOSE / GOOSE-Ex data and the PP-LiteSeg architecture. GOOSE is published under CC BY-SA 4.0. These fine-tuned checkpoints are provided under CC BY-SA 4.0; retain this notice and cite the original GOOSE / GOOSE-Ex and PP-LiteSeg publications when redistributing or using them.

The release contains no dataset images, labels or raw ROS bags.

Assets 4
Loading