GraSP instrument pipeline: weights
Weights for the GraSP surgical-instrument pipeline (boxes in, instance masks and classes out, with an uncertainty score and a SAM2 temporal correction for uncertain instances). Code, usage and results: https://github.com/volcanictv/grasp-instrument-classifier (see its README and WEIGHTS.md). The file names, sizes and sha256 checksums are in the code repository's weights_manifest.json, and python -m grasp_pipeline.weights --fetch downloads and verifies them.
Contents
| path | what |
|---|---|
weights/evidential_armN_seed{42,43,44}/ |
four-member evidential classifier ensemble per seed (final configuration, tracker-style-crop training). Seed 44 is the default. |
weights/evidential_baseline_seed{42,43,44}/ |
the same members trained without the tracker-style-crop augmentation (for reproducing the ladder) |
weights/sam2_delta.pt |
fine-tuned SAM2.1-large tensors, a delta to apply over Meta's public SAM2.1 Hiera-large checkpoint |
weights/sam3_delta.pt |
fine-tuned SAM3 mask-decoder and prompt-encoder tensors, a delta to apply over facebook/sam3 |
weights/sam{2,3}_delta_fold2.pt |
segmenter deltas of the earlier ladder rungs (fold2 cases only) |
The SAM2.1 and SAM3 base weights are not included. Get SAM2.1 from Meta (the manifest has the URL) and SAM3 from facebook/sam3 after accepting its licence.
Licence and terms
The deltas are derived from Meta's SAM2.1 (Apache 2.0) and SAM3 (gated, SAM License); use of the SAM3 delta is subject to the SAM License terms. The classifiers were trained on the GraSP dataset, which has its own terms. Check both before use. No licence has been chosen for the classifier weights yet.
Limits
Trained and evaluated on GraSP only (7 instrument classes). The pipeline takes boxes as input; reported results use ground-truth boxes. See the code repository for the measured results and their caveats.