---
license: apache-2.0
task_categories:
- feature-extraction
language:
- en
tags:
- Out-of-Distribution Detection
- Multimodal Learning
pretty_name: MultiOOD
size_categories:
- 100K<n<1M
---
<div align="center">

<h1>MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities</h1>

<div>
    <a href='https://sites.google.com/view/dong-hao/' target='_blank'>Hao Dong</a><sup>1</sup>
    &emsp;
    <a href='https://viterbi-web.usc.edu/~yzhao010/' target='_blank'>Yue Zhao</a><sup>2</sup>
    &emsp;
    <a href='https://chatzi.ibk.ethz.ch/about-us/people/prof-dr-eleni-chatzi.html' target='_blank'>Eleni Chatzi</a><sup>1</sup>
    &emsp;
    <a href='https://people.epfl.ch/olga.fink?lang=en' target='_blank'>Olga Fink</a><sup>3</sup>
</div>
<div>
    <sup>1</sup>ETH Zurich, <sup>2</sup>University of Southern California, <sup>3</sup>EPFL
</div>


<div>
    <h4 align="center">
        • <a href="https://arxiv.org/abs/2405.17419" target='_blank'>arXiv</a> •
    </h4>
</div>



<div style="text-align:center">
<img src="multiood.jpg"  width="100%" height="100%">
</div>

---

</div>

MultiOOD is the first-of-its-kind benchmark for Multimodal OOD Detection, characterized by diverse dataset sizes and varying modality combinations.

## Code
https://github.com/donghao51/MultiOOD

## MultiOOD Benchmark
MultiOOD is based on five public action recognition datasets (HMDB51, UCF101, EPIC-Kitchens, HAC, and Kinetics-600). 

### Prepare Datasets
1. Download HMDB51 video data from [link](https://serre-lab.clps.brown.edu/resource/hmdb-a-large-human-motion-database/#Downloads) and extract. Download HMDB51 optical flow data from [link](https://huggingface.co/datasets/hdong51/MultiOOD/blob/main/hmdb51_flow_mp4.tar.gz) and extract. The directory structure should be modified to match:

```
HMDB51
├── video
|   ├── catch
|   |   ├── *.avi
|   ├── climb
|   |   ├── *.avi
|   |── ...


├── flow
|   ├── *_flow_x.mp4
|   ├── *_flow_y.mp4
|   ├── ...
```

2. Download UCF101 video data from [link](https://www.crcv.ucf.edu/data/UCF101/UCF101.rar) and extract. Download UCF101 optical flow data from [link](https://huggingface.co/datasets/hdong51/MultiOOD/blob/main/ucf101_flow_mp4.tar.gz) and extract. The directory structure should be modified to match:

```
UCF101
├── video
|   ├── *.avi
|   |── ...


├── flow
|   ├── *_flow_x.mp4
|   ├── *_flow_y.mp4
|   ├── ...
```

3. Download EPIC-Kitchens video and optical flow data by 
```
bash utils/download_epic_script.sh 
```
Download audio data from [link](https://polybox.ethz.ch/index.php/s/PE2zIL99OWXQfMu).

Unzip all files and the directory structure should be modified to match:

```
EPIC-KITCHENS
├── rgb
|   ├── train
|   |   ├── D3
|   |   |   ├── P22_01.wav
|   |   |   ├── P22_01
|   |   |   |     ├── frame_0000000000.jpg
|   |   |   |     ├── ...
|   |   |   ├── P22_02
|   |   |   ├── ...
|   ├── test
|   |   ├── D3


├── flow
|   ├── train
|   |   ├── D3
|   |   |   ├── P22_01
|   |   |   |     ├── frame_0000000000.jpg
|   |   |   |     ├── ...
|   |   |   ├── P22_02
|   |   |   ├── ...
|   ├── test
|   |   ├── D3
```

4. Download HAC video, audio and optical flow data from [link](https://polybox.ethz.ch/index.php/s/3F8ZWanMMVjKwJK) and extract. The directory structure should be modified to match:

```
HAC
├── human
|   ├── videos
|   |   ├── ...
|   ├── flow
|   |   ├── ...
|   ├── audio
|   |   ├── ...

├── animal
|   ├── videos
|   |   ├── ...
|   ├── flow
|   |   ├── ...
|   ├── audio
|   |   ├── ...

├── cartoon
|   ├── videos
|   |   ├── ...
|   ├── flow
|   |   ├── ...
|   ├── audio
|   |   ├── ...
```
5. Download Kinetics-600 video data by 
```
wget -i utils/filtered_k600_train_path.txt
```
Extract all files and get audio data from video data by
```
python utils/generate_audio_files.py
```
Download Kinetics-600 optical flow data (kinetics600_flow_mp4_part_*) from [link](https://huggingface.co/datasets/hdong51/MultiOOD/tree/main) and extract (run `cat kinetics600_flow_mp4_part_* > kinetics600_flow_mp4.tar.gz` and then `tar -zxvf kinetics600_flow_mp4.tar.gz`).

Unzip all files and the directory structure should be modified to match:

```
Kinetics-600
├── video
|   ├── acting in play
|   |   ├── *.mp4
|   |   ├── *.wav
|   |── ...


├── flow
|   ├── acting in play
|   |   ├── *_flow_x.mp4
|   |   ├── *_flow_y.mp4
|   ├── ...
```

### Dataset Splits
The splits for Multimodal Near-OOD and Far-OOD Benchmarks are provided in https://github.com/donghao51/MultiOOD under `HMDB-rgb-flow/splits/` for HMDB51, UCF101, HAC, and Kinetics-600, and under `EPIC-rgb-flow/splits/` for EPIC-Kitchens.


## Methodology
<div style="text-align:left">
<img src="frame.jpg"  width="70%" height="100%">
</div>

---

An overview of the proposed framework for Multimodal OOD Detection. We introduce A2D algorithm to encourage enlarging the prediction discrepancy across modalities. Additionally, we propose a novel outlier synthesis algorithm, NP-Mix, designed to explore broader feature spaces, which complements A2D to strengthen the OOD detection performance.


## Contact
If you have any questions, please send an email to donghaospurs@gmail.com

## Citation

If you find our work useful in your research please consider citing our paper:

```
@article{dong2024multiood,
	author   = {Hao Dong and Yue Zhao and Eleni Chatzi and Olga Fink},
	title    = {{MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities}},
    journal  = {arXiv preprint arXiv:2405.17419},
	year     = {2024},
}
```