Surojit-Utah/protosemseg-checkpoints

ProtoSemSeg — Trained Model Checkpoints (Competing Baseline)

Trained checkpoints for ProtoSemSeg, a prototype-based few-shot semantic segmentation method for seismic facies, used as a competing baseline in:

Saha, S. and Whitaker, R. AdaSemSeg: An Adaptive Few-shot Semantic Segmentation of Seismic Facies. IEEE Transactions on Geoscience and Remote Sensing, 2025. arXiv:2501.16760

Code: github.com/Surojit-Utah/AdaSemSeg (see methods/protosemseg/)

This is the competing/baseline method, not the paper's proposed method. The proposed method, AdaSemSeg, and its checkpoints are at Surojit-Utah/adasemseg-checkpoints; it outperforms ProtoSemSeg on every dataset/shot combination reported in the paper (see table below). These checkpoints are provided for direct, reproducible comparison — so anyone evaluating AdaSemSeg against a class-flexible few-shot baseline doesn't have to retrain ProtoSemSeg from scratch.

What ProtoSemSeg is

ProtoSemSeg is a prototype-based few-shot segmentation method, adapted to handle a variable number of target classes across datasets — the same underlying limitation of standard few-shot semantic segmentation (FSSS) methods that motivates AdaSemSeg. It is the closest prior method identified in the AdaSemSeg paper's related work with this property, which is why it is used as the primary point of comparison rather than a generic fixed-class-count FSSS baseline.

Unlike AdaSemSeg, ProtoSemSeg does not use Gaussian Process regression or a class-agnostic binary-task decomposition with a shared backbone across all classes — see the paper (Section on competing methods) and methods/protosemseg/ for the architectural differences.

Important: what each checkpoint was trained on

Same leave-one-dataset-out protocol as AdaSemSeg — the folder name is the held-out target dataset, not the training data:

Checkpoint path	Trained (meta-training) on	Evaluated (meta-testing, no fine-tuning) on
f3/{1,5}-shot/bestmodel.pth.tar	Parihaka + Penobscot	F3
parihaka/{1,5}-shot/bestmodel.pth.tar	F3 + Penobscot	Parihaka
penobscot/{1,5}-shot/bestmodel.pth.tar	F3 + Parihaka	Penobscot

Reported performance vs. AdaSemSeg (paper Table III, FwF1)

Target dataset	Shots	AdaSemSeg	ProtoSemSeg (this repo)	Transfer learning
F3	1	0.85	0.55	0.84
F3	5	0.89	0.68	0.84
Parihaka	1	0.84	0.52	0.54
Parihaka	5	0.86	0.58	0.62
Penobscot	1	0.93	0.58	0.67
Penobscot	5	0.96	0.71	0.89

Full PA / MCA / FwIoU / FwF1 breakdowns are in the main repo README and REPRODUCE.md.

Usage

git clone https://github.com/Surojit-Utah/AdaSemSeg
cd AdaSemSeg
python scripts/download_assets.py --data
python scripts/reproduce_table3.py   # reproduces the AdaSemSeg vs. ProtoSemSeg vs. transfer-learning comparison

See methods/protosemseg/ in the main repo for the model definition and evaluation code.

Citation

If you use these checkpoints, please cite the AdaSemSeg paper (which reports this comparison) and the original ProtoSemSeg method referenced therein:

@article{saha2025adasemseg,
  title={AdaSemSeg: An Adaptive Few-shot Semantic Segmentation of Seismic Facies},
  author={Saha, Surojit and Whitaker, Ross},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2025},
  doi={10.1109/TGRS.2025.3595010}
}

@article{zhao2023protosemseg,
  author={Zhao, Yunhe and Chai, Bianfang and Shuo, Liangxun and Li, Zenghao and Wu, Heng and Wang, Tianyi},
  journal={GEOPHYSICS},
  title={Few-shot learning for seismic facies segmentation via prototype learning},
  year={2023},
  volume={88},
  number={3},
  pages={41--49}
}