HSP-IIT/fsos-ar-benchmark

FSOS-AR Benchmark Splits

Class-disjoint few-shot open-set action recognition splits used in A Baseline Study and Benchmark for Few-Shot Open-Set Action Recognition with Feature Residual Discrimination (Berti, Pasquale, Natale — accepted at ICPR 2026). Code and training scripts: hsp-iit/fsosar.

What this is (and isn't)

This repository contains only the split protocol — CSV files mapping each (class, video_id) pair to a train / validation / test partition, plus the class-name definitions where available. It does not contain the underlying video or image data. The train/validation/test classes are disjoint (no class overlap across splits), as required for few-shot evaluation: validation and test classes are never seen during training.

NTU RGB+D 120 is not included here. The paper reports results on 5 datasets, but the nturgbd120 split config and its corresponding checkpoints have been withdrawn from public release. NTU RGB+D 120's Release Agreement with ROSE Lab restricts redistribution/derivation of data from the dataset without express permission, and out of caution we removed this config (and made the two NTU RGB+D 120-trained checkpoints private) pending written confirmation from ROSE Lab. See hsp-iit/workbook-stefano-berti#70 for the full audit and status.

To reproduce the benchmark you still need to download each source dataset yourself, under its own license:

Config	Source dataset	Source license	Get it from
diving48	Diving48	No formal license published by SVCL/UCSD; research use	http://www.svcl.ucsd.edu/projects/resound/dataset.html
hmdb51	HMDB51	CC BY-NC-SA 3.0 — Non-Commercial	https://serre-lab.clps.brown.edu/resource/hmdb-a-large-human-motion-database/
ssv2	Something-Something V2	Qualcomm Data License Agreement — Research Use	https://developer.qualcomm.com/software/ai-datasets/something-something
ucf101	UCF101	No formal license published by CRCV; research use	https://www.crcv.ucf.edu/data/UCF101.php

⚠️ hmdb51 config: HMDB51's own license is CC BY-NC-SA 3.0 (non-commercial). Use of this split config — and of any checkpoints trained on it — should be limited to non-commercial research purposes, consistent with the source license, regardless of the BSD-3-Clause license on the split/label metadata itself.

The hsp-iit/fsosar GitHub repo includes helper scripts (under data/) to extract per-video frames into the folder layout expected by the training code once you have the raw videos.

Usage

from datasets import load_dataset

ds = load_dataset("HSP-IIT/fsos-ar-benchmark", "hmdb51")
print(ds["train"][0])  # {'class': 'brush_hair', 'video_id': 'April_09_brush_hair_u_nm_np1_ba_goo_0'}

Swap "hmdb51" for any of diving48, ssv2, ucf101.

hmdb51 and ucf101 additionally ship a classes.json with the natural-language class descriptions used for the textual/CLIP-based baselines in the paper.

Splits provenance

• hmdb51 / ucf101: class splits follow the ARN protocol.
• ssv2: class splits follow the OTAM protocol (class folders are the anonymized identifiers used by that protocol, not literal class names).
• diving48: splits introduced in this paper, built with data/create_train_test_diving_ntu.py in the code repo.

Citation

@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}
}

License

The split files and class definitions in this repository are released under the BSD 3-Clause License (matching the code repository). This license covers only the split/label metadata created by the authors, not the underlying video datasets listed above, which remain subject to their original licenses. In particular:

• HMDB51 (CC BY-NC-SA 3.0): non-commercial use only.
• Something-Something V2: Qualcomm's Data License Agreement restricts use to research purposes.

NTU RGB+D 120 splits/checkpoints have been withdrawn from this release; see the note above.

A full license audit of this project (code, paper, splits, checkpoints, and dependencies) is tracked at hsp-iit/workbook-stefano-berti#70.