File size: 6,068 Bytes
5193ee6 57f7933 5193ee6 57f7933 5193ee6 57f7933 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 |
---
base_model:
- Qwen/Qwen2-VL-7B-Instruct
language:
- en
license: apache-2.0
metrics:
- accuracy
pipeline_tag: image-text-to-text
library_name: transformers
---
# DeepPerception: Advancing R1-like Cognitive Visual Perception in MLLMs for Knowledge-Intensive Visual Grounding
This is the official repository of **DeepPerception**, an MLLM enhanced with cognitive visual perception capabilities.
[Project Page](https://deepperception-kvg.github.io/)
[Paper](https://arxiv.org/abs/2503.12797)
## Overview
<p align="center">
<img src="figs/header.png" width="100%"></a><br>
Figure 1: (a) <strong>DeepPerception</strong> employs knowledge-driven reasoning to derive answers, while the baseline model directly outputs predictions without cognitive processing. (b) <strong>DeepPerception</strong> demonstrates superior cognitive visual perception capabilities that cannot be elicited in the foundation model through simplistic zero-shot CoT prompting.
</p>
#### Abstract
Human experts excel at fine-grained visual discrimination by leveraging domain knowledge to refine perceptual features, a capability that remains underdeveloped in current Multimodal Large Language Models (MLLMs). Despite possessing vast expert-level knowledge, MLLMs struggle to integrate reasoning into visual perception, often generating direct responses without deeper analysis.
To bridge this gap, we introduce knowledge-intensive visual grounding (KVG), a novel visual grounding task that requires both finegrained perception and domain-specific knowledge integration. To address the challenges of KVG, we propose **DeepPerception**, an MLLM enhanced with cognitive visual perception capabilities. Our approach consists of (1) an automated data synthesis pipeline that generates high-quality, knowledge-aligned training samples, and (2) a two-stage training framework combining supervised fine-tuning for cognitive reasoning scaffolding and reinforcement learning to optimize perceptioncognition synergy. To benchmark performance, we introduce KVG-Bench, a comprehensive dataset spanning 10 domains with 1.3K manually curated test cases.
Experimental results demonstrate that DeepPerception significantly outperforms direct fine-tuning, achieving +8.08% accuracy improvements on KVG-Bench and exhibiting +4.60% superior cross-domain generalization over baseline approaches. Our findings highlight the importance of integrating cognitive processes into MLLMs for human-like visual perception and open new directions for multimodal reasoning research.
#### Key Contributions
- We introduce the task of **Knowledge-intensive Visual Grounding (KVG)** to explore the concept of cognitive visual perception for MLLMs, aiming to integrate their inherent knowledge and reasoning capabilities into visual perception.
- We propose **[DeepPerception](https://huggingface.co/MaxyLee/DeepPerception)**, an MLLM with enhanced cognitive visual perception capabilities. To achieve this, we develop an automated dataset creation pipeline and a two-stage framework integrating supervised cognitive capability enhancement with perception-oriented reinforcement learning.
- We introduce **[KVG-Bench](https://huggingface.co/datasets/MaxyLee/KVG-Bench)**, a manually curated benchmark for the KVG task involving diverse knowledge domains and entities. Experiments on KVG-Bench and other fine-grained visual recognition tasks demonstrate DeepPerception's exceptional cognitive visual perception capabilities and superior cross-domain generalization performance.
## Get Started
### Contents:
- [Environment](#environment)
- [Data Preparation](#data-preparation)
- [Checkpoints](#checkpoints)
- [Evaluation](#evaluation)
- [Training](#training)
### Environment
1. Clone this repository and navigate to DeepPerception folder
```bash
git clone https://github.com/MaxyLee/DeepPerception.git
cd DeepPerception
```
2. Install Packages
For evaluation:
```bash
conda env create -n deepperception python=3.9
conda activate deepperception
pip install -r requirements.txt
```
### Data Preparation
| Dataset | Links |
|--------- |---------------------------------------|
| KVG-Bench | [`🤗HuggingFace`](https://huggingface.co/datasets/MaxyLee/KVG-Bench) |
| KVG Training | [`🤗HuggingFace`](https://huggingface.co/datasets/MaxyLee/KVG) |
---
### Checkpoints
| Model | Links |
|--------- |---------------------------------------|
| DeepPerception | [`🤗HuggingFace`](https://huggingface.co/MaxyLee/DeepPerception) |
| DeepPerception-FGVR | [`🤗HuggingFace`](https://huggingface.co/MaxyLee/DeepPerception-FGVR) |
---
### Evaluation
```bash
# Evaluate on KVG-Bench
bash eval.sh [CUDA_IDS] [KVG_BENCH_PATH] [CKPT_PATH]
```
Notice: Please modify the script if you want to evaluate on Qwen2-VL.
### Training
TODO
## Citation
If you find DeepPerception useful for your research or applications, please cite using this BibTeX:
```bibtex
@misc{ma2025deepperception,
title={DeepPerception: Advancing R1-like Cognitive Visual Perception in MLLMs for Knowledge-Intensive Visual Grounding},
author={Xinyu Ma and Ziyang Ding and Zhicong Luo and Chi Chen and Zonghao Guo and Derek F. Wong and Xiaoyi Feng and Maosong Sun},
year={2025},
url={https://arxiv.org/abs/2503.12797},
}
```
## Acknowledgement
- [Qwen2-VL](https://github.com/QwenLM/Qwen2.5-VL)
- [vLLM](https://github.com/vllm-project/vllm)
- [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory)
- [R1-V](https://github.com/Deep-Agent/R1-V)
## License
[](https://github.com/twbs/bootstrap/blob/main/LICENSE)
[](https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE) |