Rowance/HSA-features

HSA Frozen Feature Archives

Paper · Code · Upstream sources

Frozen ImageBind and LanguageBind features for the main retrieval and prototype-classification experiments of Hub-Spectral Activation of Latent Multimodal Knowledge. With these archives, the HSA and frozen-cosine results can be recomputed with the public code without downloading the original media or encoder weights.

This repository contains 26 of the 30 main-experiment archives (17 retrieval and 9 classification relations, 1,736,396,029 bytes). The four Ego4D archives are not included: Ego4D data is available only under the Ego4D license, and licensees can generate these archives with download.py and extract.py in the code repository.

Contents

Source dataset	Encoder	Retrieval	Classification	Archive terms
TartanRGBT	LanguageBind	4	3	BSD-3-Clause-Clear
TartanRGBT	ImageBind	2	1	CC BY-NC-SA 4.0
VGGSound	LanguageBind	1	0	CC BY 4.0
VGGSound / VGGSound-309	ImageBind	1	1	CC BY-NC-SA 4.0
Caltech Aerial RGBT	ImageBind	1	0	CC BY-NC-SA 4.0
BatVision	LanguageBind	1	0	CC BY-SA 4.0
BatVision	ImageBind	1	1	CC BY-SA 4.0
UTD-MHAD	ImageBind	1	1	CC BY-NC-SA 4.0
UCF101	LanguageBind	1	0	CC BY-NC 4.0
NYUv2	LanguageBind	1	1	CC BY-NC 4.0
MSR-VTT	LanguageBind	1	1	CC BY-NC 4.0
MAVD	ImageBind	1	0	MAVD license (non-commercial)
MAVD	LanguageBind	1	0	MAVD license (non-commercial)
Total		17	9

The terms of each file are also recorded as proposed_feature_release_terms in manifest.json.

• The ImageBind license (CC BY-NC-SA 4.0) does not state whether it covers features computed with the model. ImageBind-derived archives are released under CC BY-NC-SA 4.0 except where the source dataset requires its own terms (BatVision: CC BY-SA 4.0; MAVD: MAVD license).
• UTD-MHAD, UCF101, NYUv2 and MSR-VTT publish no license. Their archives are released for non-commercial research use under the terms in the table; use of the underlying data remains subject to the original datasets' terms.
• MAVD notice. The MAVD-derived archives are modified versions of MAVD data: they contain feature vectors computed from MAVD RGB, thermal and audio data (and scene-description text) by frozen encoders, assembled on 2026-09-16. They are distributed for non-commercial purposes only under the MAVD license (LICENSES/MAVD_RLLA.txt), including its disclaimer of warranty and limitation of liability.

If you hold rights in a source dataset and object to an archive, please open an issue in the code repository and we will remove it. Please cite the original datasets and encoders you use (see Citation).

Files

retrieval/{imagebind,languagebind}/<relation>.npz
classification/{imagebind,languagebind}/<relation>.npz
manifest.json            # shapes, sizes, SHA-256, dataset, protocol hash and terms per archive
SHA256SUMS               # checksums of the 26 archives
reference_results.json   # per-relation HSA / frozen-cosine results on these archives
protocols/               # sample IDs, splits, prompts and sources at the evaluation commit
UPSTREAM_SOURCES.md      # dataset and encoder sources and terms
LICENSES/                # ImageBind, LanguageBind, VGGSound and MAVD license texts

Array format

Load with numpy.load(path, allow_pickle=False). N is the training count, M the test count, D the target feature dimension, H the hub dimension and C the number of classes.

Array	Shape	Description
a_train, h_a_train	(N, D), (N, H)	Modality A and hub features, paired row by row
b_train, h_b_train	(N, D), (N, H)	Modality B and hub features, paired row by row
a_test, b_test	(M, D) each	Retrieval queries and candidates
positive_labels	(M,)	Optional; a candidate is positive when its label matches the query label, otherwise when its row index matches
prototypes	(C, D)	Classification prototype bank in modality A
queries	(M, D)	Classification queries in modality B
labels	(M,)	Integer prototype index of each query

The registered protocols use the same training samples for both hub edges, so h_a_train and h_b_train contain the same observations. Sample IDs, split order, prompts and preprocessing are fixed in protocols/ and in the code.

Usage

git clone https://github.com/Luo1Yan/HSA.git
cd HSA
git checkout 64b3e9ca01abefb878b02f546f0b593ee459d6cb
python -m pip install -r requirements.txt huggingface_hub
hf download Rowance/HSA-features --repo-type dataset --local-dir datasets
for d in tartan-rgbt vggsound caltech-rgbt batvision utd-mhad ucf101 nyuv2 msrvtt mavd; do
  python run.py --dataset $d --output outputs/$d.json
done

Select by dataset as above: running python run.py without a selection expects all 30 archives, including the four Ego4D archives. Add --device cuda to score retrieval on a GPU.

Equal-weight means over the relations in this repository:

Task	Relations	Frozen cosine	HSA
Bidirectional retrieval R@10	17	20.23%	33.77%
Classification macro Top-1	9	33.93%	59.14%

The paper's means over all 30 relations are 18.27% / 31.15% (retrieval R@10, frozen cosine / HSA) and 29.01% / 52.43% (classification macro Top-1). Per-relation values are in reference_results.json.

Provenance

The archives were extracted with frozen ImageBind-Huge (source commit 53680b02) and LanguageBind (source commit 7070c533; checkpoint revisions in protocols/downloads.json). All file sizes and SHA-256 hashes match manifest.json, and all 26 archives pass run.py --check-data at commit 64b3e9c. End-to-end re-extraction from raw data has been verified on UTD-MHAD. LanguageBind video preprocessing applies random horizontal flips; the flip state of the archived video features was not recorded, so features re-extracted from raw videos may differ slightly from these archives.

Citation

@misc{guo2026hsa,
  title={Hub-Spectral Activation of Latent Multimodal Knowledge},
  author={Guo, Ying and Chen, Haidong and Xu, Linrui and Liu, Xiaohao and
          Shi, Chuancheng and Xiao, Canran and Zhang, Dan and Shen, Fei and
          Shen, Li and Chua, Tat-Seng},
  year={2026},
  eprint={2609.17094},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.17094}
}

Please also cite the source datasets and encoders you use:

• TartanRGBT: P. Maheshwari et al., AnyThermal: Towards Learning Universal Representations for Thermal Perception, ICRA 2026 (arXiv:2602.06203).
• VGGSound: H. Chen, W. Xie, A. Vedaldi, A. Zisserman, VGGSound: A Large-Scale Audio-Visual Dataset, ICASSP 2020.
• Caltech Aerial RGBT: C. Lee et al., Caltech Aerial RGB-Thermal Dataset in the Wild, CaltechDATA (doi:10.22002/cks6g-ps927).
• BatVision: A. Brunetto, S. Hornauer, S. X. Yu, F. Moutarde, The Audio-Visual BatVision Dataset for Research on Sight and Sound, IROS 2023 (data doi:10.57745/HYLZNL).
• UTD-MHAD: C. Chen, R. Jafari, N. Kehtarnavaz, UTD-MHAD: A Multimodal Dataset for Human Action Recognition Utilizing a Depth Camera and a Wearable Inertial Sensor, ICIP 2015.
• UCF101: K. Soomro, A. R. Zamir, M. Shah, UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild, CRCV-TR-12-01, 2012 (arXiv:1212.0402).
• NYUv2: N. Silberman, D. Hoiem, P. Kohli, R. Fergus, Indoor Segmentation and Support Inference from RGBD Images, ECCV 2012.
• MSR-VTT: J. Xu, T. Mei, T. Yao, Y. Rui, MSR-VTT: A Large Video Description Dataset for Bridging Video and Language, CVPR 2016.
• MAVD: F. Rivera Valverde, J. V. Hurtado, A. Valada, There is More than Meets the Eye: Self-Supervised Multi-Object Detection and Tracking with Sound by Distilling Multimodal Knowledge, CVPR 2021.
• ImageBind: R. Girdhar et al., ImageBind: One Embedding Space To Bind Them All, CVPR 2023.
• LanguageBind: B. Zhu et al., LanguageBind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment, ICLR 2024.