|
--- |
|
license: apache-2.0 |
|
--- |
|
# R1-Omni: Explainable Omni-Multimodal Emotion Recognition with Reinforcement Learning |
|
|
|
[](https://github.com/Jiaxing-star/R1-Omni) |
|
[](https://modelscope.cn/models/iic/R1-Omni-0.5B) |
|
[](https://arxiv.org/abs/2503.05379) |
|
|
|
|
|
## 📖 Introduction |
|
**R1-Omni** is the industry’s first application of Reinforcement Learning with Verifiable Reward (RLVR) to an Omni-multimodal large language model. We focus on emotion recognition, a task where both visual and audio modalities play crucial roles, to validate the potential of combining RLVR with Omni model. Our findings reveal several key insights: |
|
1) **Enhanced Reasoning Capability**: R1-Omni demonstrate 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. |
|
|
|
|
|
## 🏆 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**. To ensure the model runs inference smoothly, follow these steps: |
|
|
|
1. **Download the Required Models**: |
|
- [siglip-224](https://huggingface.co/google/siglip-base-patch16-224). |
|
- [whisper-large-v3](https://huggingface.co/openai/whisper-large-v3). |
|
|
|
2. **Update the Configuration File**: |
|
- In the directory where you downloaded the R1-Omni model, locate the config.json file. |
|
- Update the paths on line 23 and line 31 to point to the local folders where you saved the models. |
|
|
|
|
|
#### Example: Updating config.json |
|
If you saved the models to the following local paths:: |
|
- `/path/to/local/models/siglip-base-patch16-224` |
|
- `/path/to/local/models/whisper-large-v3` |
|
|
|
Update the relevant lines in config.json as follows: |
|
```json |
|
"mm_audio_tower": "/path/to/local/models/whisper-large-v3", |
|
"mm_vision_tower": "/path/to/local/models/siglip-base-patch16-224" |
|
``` |
|
|
|
We provide inference.py for singe video inference. |
|
``` |
|
python inference.py --modal video_audio \ |
|
--model_path ./R1-Omni-0.5B \ |
|
--video_path video.mp4 \ |
|
--instruct "As an emotional recognition expert; throughout the video, which emotion conveyed by the characters is the most obvious to you? Output the thinking process in <think> </think> and final emotion in <answer> </answer> tags." |
|
``` |
|
|
|
|
|
|
|
## 🧠 Training |
|
### Cold Start |
|
we initialize the HumanOmni-0.5B by fine-tuning it on a combined dataset consisting of 232 samples from the [Explainable Multimodal Emotion Reasoning](https://github.com/zeroQiaoba/AffectGPT) dataset and 348 samples from HumanOmni dataset. |
|
An example json file of the training data: |
|
``` |
|
[ |
|
{ |
|
"video": "MER24/sample_00000967.mp4", |
|
"conversations": [ |
|
{ |
|
"from": "human", |
|
"value": "<video>\n<audio>\nAs an emotional recognition expert; throughout the video, which emotion conveyed by the characters is the most obvious to you? Output the thinking process in <think> </think> and final emotion in <answer> </answer> tags." |
|
}, |
|
{ |
|
"from": "gpt", |
|
"value": "<think>The video depicts a bright and tranquil indoor setting, where a man in a white Polo shirt stands by the window, engaged in a phone call. His furrowed brow and open mouth suggest he is experiencing tension and anxiety. According to the audio content of the video, his speech is fast-paced, and his tone is filled with confusion and stress. A comprehensive analysis reveals that the man is facing a moderate level of anxiety, closely linked to the challenging phone conversation he is having. Consequently, the entire emotional analysis report emphasizes his anxiety and nervousness in handling challenging situations.</think>\n<answer>anxious</answer>" |
|
} |
|
] |
|
}, |
|
... |
|
] |
|
``` |
|
All of the cold-start data will be released as soon as possible. |
|
|
|
### RLVR |
|
In this stage, we utilize the training sets from the [MAFW](https://mafw-database.github.io/MAFW/) and DFEW(https://dfew-dataset.github.io/) datasets, comprising a total of 15,306 video sample. |
|
An example json file of the training data: |
|
``` |
|
[ |
|
{ |
|
"video": "DFEW/videos/1.mp4", |
|
"conversations": [ |
|
{ |
|
"from": "human", |
|
"value": "<video>\n<audio>\nAs an emotional recognition expert; throughout the video, which emotion conveyed by the characters is the most obvious to you?" |
|
}, |
|
{ |
|
"from": "gpt", |
|
"value": "sad" |
|
} |
|
], |
|
}, |
|
... |
|
] |
|
``` |
|
|
|
|
|
## 🤝 Related Work |
|
- [R1-V](https://github.com/Deep-Agent/R1-V) |
|
- [HumanOmni](https://github.com/HumanMLLM/HumanOmni) |
|
- [DeepSeek-R1](https://github.com/deepseek-ai/DeepSeek-R1) |
|
|
|
## 📚 Citation |
|
If you find our work helpful, feel free to give us a cite. |
|
``` |
|
{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} |
|
} |
|
``` |