|
--- |
|
license: cc-by-nc-4.0 |
|
library_name: transformers |
|
pipeline_tag: video-text-to-text |
|
--- |
|
|
|
# R1-Omni: Explainable Omni-Multimodal Emotion Recognition with Reinforcement Learning |
|
|
|
[](https://modelscope.cn/models/iic/R1-Omni-0.5B) |
|
[](https://huggingface.co/StarJiaxing/R1-Omni-0.5B) |
|
[](https://arxiv.org/abs/2503.05379) |
|
|
|
This model utilizes Reinforcement Learning with Verifiable Reward (RLVR) to perform omni-multimodal emotion recognition. Built upon the HumanOmni-0.5B model, R1-Omni excels at understanding visual and audio cues for emotion identification, even in out-of-distribution scenarios. |
|
|
|
## 📖 Introduction |
|
**R1-Omni** is the first application of Reinforcement Learning with Verifiable Reward (RLVR) to an Omni-multimodal large language model. It focuses on emotion recognition, where visual and audio modalities play crucial roles. Key insights include: |
|
|
|
1) **Enhanced Reasoning Capability**: R1-Omni demonstrates superior reasoning abilities, enabling a clearer understanding of how visual and audio information contribute to emotion recognition. |
|
2) **Improved Understanding Capability**: Compared to SFT, RLVR significantly boosts performance on emotion recognition tasks. |
|
3) **Stronger Generalization Capability**: RLVR models exhibit markedly better generalization capabilities, particularly excelling in out-of-distribution scenarios. |
|
|
|
## 📦 Model Download |
|
The model is based on the open-source HumanOmni-0.5B model. The following models are available: HumanOmni-0.5B, the cold-start model EMER-SFT, the model MAFW-DFEW-SFT fine-tuned directly on the MAFW and DFEW training sets, and the final model R1-Omni. |
|
|
|
<div align="center"> |
|
|
|
| **Model** | **HuggingFace** | **ModelScope** | |
|
|------------------------|---------------------------------------------------------------------------------|-------------------------------------------------------------------------| |
|
| `HumanOmni-0.5B` | [](https://hf.co/StarJiaxing/HumanOmni-0.5B) | [](https://modelscope.cn/models/iic/HumanOmni-0.5B) | |
|
| `EMER-SFT` | [](https://hf.co/StarJiaxing/EMER-SFT-0.5B) | [](https://modelscope.cn/models/iic/EMER-SFT-0.5B) | |
|
| `MAFW-DFEW-SFT` | [](https://huggingface.co/StarJiaxing/MAFW-DFEW-0.5B) | [](https://modelscope.cn/models/iic/MAFW-DFEW-0.5B) | |
|
| `R1-Omni` | [](https://huggingface.co/StarJiaxing/R1-Omni-0.5B) | [](https://modelscope.cn/models/iic/R1-Omni-0.5B) | |
|
</div> |
|
|
|
## 🏆 Performance |
|
|
|
Below are the performance on emotion recognition datasets. We use symbols to indicate whether the data is **in-distribution (⬤)** or **out-of-distribution (△)**. |
|
|
|
| Method | DFEW (WAR) ⬤ | DFEW (UAR) ⬤ | MAFW (WAR) ⬤ | MAFW (UAR) ⬤ | RAVDESS (WAR) △ | RAVDESS (UAR) △ | |
|
|----------------------------------|---------------|---------------|---------------|---------------|------------------|------------------| |
|
| HumanOmni-0.5B | 22.64 | 19.44 | 20.18 | 13.52 | 7.33 | 9.38 | |
|
| EMER-SFT | 38.66 | 35.31 | 38.39 | 28.02 | 29.00 | 27.19 | |
|
| MAFW-DFEW-SFT | 60.23 | 44.39 | 50.44 | 30.39 | 29.33 | 30.75 | |
|
| R1-Omni | 65.83 | 56.27 | 57.68 | 40.04 | 43.00 | 44.69 | |
|
|
|
 |
|
|
|
### Legend |
|
- **⬤**: Indicates **in-distribution data** (DFEW and MAFW). |
|
- **△**: Indicates **out-of-distribution data** (RAVDESS). |
|
|
|
## 🛠️ Environment Setup |
|
Our code is built on the R1-V framework. To set up the environment, please follow the installation instructions in the [R1-V repository](https://github.com/Deep-Agent/R1-V/) |
|
|
|
## 🔍 Inference |
|
Our inference code is based on the implementation from **HumanOmni**. |
|
|
|
## 📚 Citation |
|
If you find our work helpful, feel free to cite us. |
|
``` |
|
{zhao2025r1omniexplainableomnimultimodalemotion, |
|
title={R1-Omni: Explainable Omni-Multimodal Emotion Recognition with Reinforcement Learning}, |
|
author={Jiaxing Zhao and Xihan Wei and Liefeng Bo}, |
|
journal={arXiv preprint arXiv:2503.05379}, |
|
year={2025} |
|
} |
|
``` |
