--- license: bsd-3-clause tags: - action-recognition - few-shot-learning - open-set-recognition - video - pytorch library_name: pytorch --- # SAFSAR — FR-Disc — UCF101 (5-shot) Checkpoint from [*A Baseline Study and Benchmark for Few-Shot Open-Set Action Recognition with Feature Residual Discrimination*](https://arxiv.org/abs/2603.04125) (Berti, Pasquale, Natale — ICPR 2026), trained with the **discriminator (FR-Disc)** open-set loss. - Model: `SAFSAR` - Dataset: UCF101 - Setting: 5-way 5-shot - Open-set method: FR-Disc (feature-residual discriminator) - Checkpoint eval metrics (measured for this exact checkpoint): FS-ACC = 98.83, OS-ACC = 94.95 The paper's Table reports **99.28 FS-ACC / 91.52 OS-ACC** for this setting (averaged over evaluation episodes with a different random seed than this specific checkpoint's saved eval numbers below); the two should be close but are not guaranteed to match exactly. Code, other checkpoints, and the [FSOS-AR benchmark splits](https://huggingface.co/datasets/HSP-IIT/fsos-ar-benchmark) are at [hsp-iit/fsosar](https://github.com/hsp-iit/fsosar). ## Usage ```python from huggingface_hub import hf_hub_download import torch ckpt_path = hf_hub_download(repo_id="HSP-IIT/fsosar-safsar-frdisc-ucf101-5shot", filename="model.pt") state_dict = torch.load(ckpt_path, map_location="cpu") # Requires the model definition from https://github.com/hsp-iit/fsosar from models import SAFSAR from utils import load_configs config = load_configs("SAFSAR", "UCF101") model = SAFSAR(config, disc=True, gc=False, dp=config["dp"]) model.load_state_dict(state_dict, strict=True) model.eval() ``` ## License BSD 3-Clause, matching the [code repository](https://github.com/hsp-iit/fsosar). ## Citation ```bibtex @inproceedings{berti2026fsosar, title = {A Baseline Study and Benchmark for Few-Shot Open-Set Action Recognition with Feature Residual Discrimination}, author = {Berti, Stefano and Pasquale, Giulia and Natale, Lorenzo}, booktitle = {International Conference on Pattern Recognition (ICPR)}, year = {2026} } ```