---
pretty_name: HSA Frozen Features
license:
  - cc-by-nc-sa-4.0
  - cc-by-nc-4.0
  - cc-by-sa-4.0
  - cc-by-4.0
  - bsd-3-clause-clear
  - other
task_categories:
  - feature-extraction
tags:
  - multimodal
  - cross-modal-retrieval
  - imagebind
  - languagebind
viewer: false
---

# HSA Frozen Features

[Paper](https://arxiv.org/abs/2609.17094) · [Code](https://github.com/Luo1Yan/HSA)

Precomputed ImageBind and LanguageBind features for the main experiments of
[*Hub-Spectral Activation of Latent Multimodal Knowledge*](https://arxiv.org/abs/2609.17094).
With them you can reproduce the paper's HSA and frozen-cosine results without
downloading the original datasets or the encoders.

## Quick start

```bash
git clone https://github.com/Luo1Yan/HSA.git
cd HSA
python -m pip install -r requirements.txt huggingface_hub
hf download Rowance/HSA-features --repo-type dataset --local-dir datasets
python run.py --dataset utd-mhad --output outputs/utd-mhad.json
```

Replace `utd-mhad` with any dataset below to evaluate it. The UTD-MHAD run
reports 9.30% retrieval R@10 and 34.75% classification macro Top-1 for HSA.
Expected results for every experiment are in `reference_results.json`.

## What is included

Features for 26 of the 30 main experiments, covering every dataset except Ego4D.

| Dataset (`--dataset`) | Retrieval | Classification | License |
|---|---:|---:|---|
| TartanRGBT (`tartan-rgbt`) | 6 | 4 | BSD-3-Clause-Clear (LanguageBind), CC BY-NC-SA 4.0 (ImageBind) |
| VGGSound (`vggsound`) | 2 | 1 | CC BY 4.0 (LanguageBind), CC BY-NC-SA 4.0 (ImageBind) |
| Caltech Aerial RGBT (`caltech-rgbt`) | 1 | 0 | CC BY-NC-SA 4.0 |
| BatVision (`batvision`) | 2 | 1 | CC BY-SA 4.0 |
| UTD-MHAD (`utd-mhad`) | 1 | 1 | CC BY-NC-SA 4.0 |
| UCF101 (`ucf101`) | 1 | 0 | CC BY-NC 4.0 |
| NYUv2 (`nyuv2`) | 1 | 1 | CC BY-NC 4.0 |
| MSR-VTT (`msrvtt`) | 1 | 1 | CC BY-NC 4.0 |
| MAVD (`mavd`) | 2 | 0 | MAVD license (non-commercial) |

Ego4D features are not included because Ego4D is distributed only under its own
license. Ego4D licensees can generate them with the code repository.

Each `.npz` file holds the feature matrices, labels and class prototypes of one
experiment; the [feature format](https://github.com/Luo1Yan/HSA/blob/main/docs/feature-format.md)
describes the arrays. LanguageBind video preprocessing includes random
horizontal flips, so video features re-extracted from the original data can
differ slightly from these files.

## License

Each file's license is listed in the table above and in `manifest.json`. Please
also follow the terms of the original datasets (see `UPSTREAM_SOURCES.md`).

The MAVD files are modified versions of MAVD data (feature vectors computed from
MAVD RGB, thermal, audio and scene-description text, 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 one of the source datasets and want a file removed, please
open an issue in the [code repository](https://github.com/Luo1Yan/HSA/issues).

## Citation

```bibtex
@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 datasets and encoders you use: TartanRGBT (Maheshwari et
al., ICRA 2026), VGGSound (Chen et al., ICASSP 2020), Caltech Aerial RGBT
(Lee et al., 2024), BatVision (Brunetto et al., IROS 2023), UTD-MHAD (Chen et
al., ICIP 2015), UCF101 (Soomro et al., 2012), NYUv2 (Silberman et al., ECCV
2012), MSR-VTT (Xu et al., CVPR 2016), MAVD (Rivera Valverde et al., CVPR
2021), ImageBind (Girdhar et al., CVPR 2023) and LanguageBind (Zhu et al.,
ICLR 2024).