KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #developer-tools #ai-agent #claude-code #codex #typescript #ai-agents #deepseek-harness #mcp #cli #rust

Long-Tail-image-aesthetics-and-quality-assessment

★ 91 Open GitHub ↗

🔥[ICML 2024, Official Code] First work to propose a solution to the long-tail problem in IAA. 首篇针对IAA中的长尾问题提出解决方案的工作

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Long-Tail-image-aesthetics-and-quality-assessment
ICML2024: First work to propose a solution to the long-tail problem in IAA. 首篇针对IAA中的长尾问题提出解决方案的工作

[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Framework](https://img.shields.io/badge/PyTorch-%23EE4C2C.svg?&logo=PyTorch&logoColor=white)](https://pytorch.org/)

[[国内的小伙伴可以看这]](https://github.com/woshidandan/Long-Tail-image-aesthetics-and-quality-assessment/blob/main/CN_README.md)This repo contains the official implementation of ELTA of the **ICML 2024** paper. Our ELTA 2.0 version released in IJCV2026 [here](https://github.com/woshidandan/Rethinking-Long-Tail-for-Image-Aesthetics-Assessment).

<div align="center">
<h1>

ELTA: An Enhancer against Long-Tail for Aesthetics-oriented Models

</h1>
<h4>

Limin Liu*, Shuai He*, Anlong Ming*, Rui Xie, Huadong Ma
 
Beijing University of Posts and Telecommunications, *Equal contribution

</h4>
</div>

-----------------------------------------

## Introduction
 Real-world datasets often exhibit long-tailed distributions, compromising the generalization and fairness of learning-based models. This issue is particularly pronounced in Image Aesthetics Assessment (IAA) tasks, where such imbalance is difficult to mitigate due to a severe distribution mismatch between features and labels, as well as the great sensitivity of aesthetics to image variations.
To address these issues, we propose an Enhancer against Long-Tail for Aesthetics-oriented models (ELTA). ELTA first utilizes a dedicated mixup technique to enhance minority feature representation in high-level space while preserving their intrinsic aesthetic qualities. Next, it aligns features and labels through a similarity consistency approach, effectively alleviating the distribution mismatch. Finally, ELTA adopts a specific strategy to refine the output distribution, thereby enhancing the quality of pseudo-labels.

## Environment Installation
* einops==0.4.1
* matplotlib==3.3.4
* nni==2.6.1
* numpy==1.19.5
* pandas==1.1.5
* Pillow==10.2.0
* scikit_learn==1.4.0
* scipy==1.5.4
* timm==0.6.12
* torch==1.10.1
* torchvision==0.11.2
* tqdm==4.64.1

## Model training
```
python main.py --csv_path           [dataset annotation file path]
               --dataset_path       [dataset image path]
               --mixup              # optional, enable TFA module
               --simloss_weight 1   # optional, enable FLSA module and specify weight
               ...                  # other arguments
```

checkpoint file url: https://drive.google.com/file/d/1pA7kOCPHEUR5oNnocBZHH41Erud9Y30S/view?usp=drive_link
The direct inference code: https://drive.google.com/file/d/1DqJnRNrwC0XpwYAKoOK3A7C_Vazc1fNs/view?usp=sharing

## Model evaluation
```
python main.py -e                   [dataset annotation file path]
               --test_dataset_path  [dataset image path]
               --resume             [checkpoint path]   # required!
               ...                  # other arguments
```

## Model self-training (after the evaluation)
```
python main.py --st                 # enable self-training
               ...                  # other arguments
```

## Recommended: use the NNI for automatic parameter tuning
```
# Modify the 'trial_command' and 'search_space' in the file 'main_nni.py'
python main_nni.py
```

## If you find our work is useful, pleaes cite our paper:
```
@inproceedings{liuelta,
  title={ELTA: An Enhancer against Long-Tail for Aesthetics-oriented Models},
  author={Liu, Limin and He, Shuai and Ming, Anlong and Xie, Rui and Ma, Huadong},
  booktitle={Forty-first International Conference on Machine Learning}
}
```

## Related Work from Our Group
<table>
 <thead align="center">
 <tr>
 <td>🎁 Projects</td>
 <td>📚 Publication</td>
 <td>🌈 Content</td>
 <td>⭐ Stars</td>
 </tr>
 </thead>
 <tbody>
 <tr>
 <td><a href="https://github.com/woshidandan/Attacker-against-image-aesthetics-assessment-model">Attacker Against IAA Model【美学模型的攻击和安全评估框架】</a></td>
 <td>TIP 2025</td>
 <td>Code, Dataset</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/Attacker-against-image-aesthetics-assessment-model?style=flat-square&labelColor=343b41"/></td>
 </tr
 <tr>
 <td><a href="https://github.com/woshidandan/Rethinking-Personalized-Aesthetics-Assessment">Personalized Aesthetics Assessment【个性化美学评估新范式】</a></td>
 <td>CVPR 2025</td>
 <td>Code, Dataset</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/Rethinking-Personalized-Aesthetics-Assessment?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/Pixel-level-No-reference-Image-Exposure-Assessment">Pixel-level image exposure assessment【首个像素级曝光评估】</a></td>
 <td>NIPS 2024</td>
 <td>Code, Dataset</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/Pixel-level-No-reference-Image-Exposure-Assessment?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/Long-Tail-image-aesthetics-and-quality-assessment">Long-tail solution for image aesthetics assessment【美学评估数据不平衡解决方案】</a></td>
 <td>ICML 2024</td>
 <td>Code</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/Long-Tail-image-aesthetics-and-quality-assessment?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/Prompt-DeT">CLIP-based image aesthetics assessment【基于CLIP多因素色彩美学评估】</a></td>
 <td>Information Fusion 2024</td>
 <td>Code, Dataset</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/Prompt-DeT?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/SR-IAA-image-aesthetics-and-quality-assessment">Compare-based image aesthetics assessment【基于对比学习的多因素美学评估】</a></td>
 <td>ACMMM 2024</td>
 <td>Code</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/SR-IAA-image-aesthetics-and-quality-assessment?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/Image-Color-Aesthetics-and-Quality-Assessment">Image color aesthetics assessment【首个色彩美学评估】</a></td>
 <td>ICCV 2023</td>
 <td>Code, Dataset</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/Image-Color-Aesthetics-and-Quality-Assessment?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/Image-Aesthetics-and-Quality-Assessment">Image aesthetics assessment【通用美学评估】</a></td>
 <td>ACMMM 2023</td>
 <td>Code</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/Image-Aesthetics-and-Quality-Assessment?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/TANet-image-aesthetics-and-quality-assessment">Theme-oriented image aesthetics assessment【首个多主题美学评估】</a></td>
 <td>IJCAI 2022</td>
 <td>Code, Dataset</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/TANet-image-aesthetics-and-quality-assessment?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/AK4Prompts">Select prompt based on image aesthetics assessment【基于美学评估的提示词筛选】</a></td>
 <td>IJCAI 2024</td>
 <td>Code</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/AK4Prompts?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/mRobotit/M2Beats">Motion rhythm synchronization with beats【动作与韵律对齐】</a></td>
 <td>IJCAI 2024</td>
 <td>Code, Dataset</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/mRobotit/M2Beats?style=flat-square&labelColor=343b41"/></td>
 </tr>
 <tr>
 <td><a href="https://github.com/woshidandan/Champion-Solution-for-CVPR-NTIRE-2024-Quality-Assessment-on-AIGC">Champion Solution for AIGC Image Quality Assessment【NTIRE AIGC图像质量评估赛道冠军】</a></td>
 <td>CVPRW NTIRE 2024</td>
 <td>Code</td>
 <td><img alt="Stars" src="https://img.shields.io/github/stars/woshidandan/Champion-Solution-for-CVPR-NTIRE-2024-Quality-Assessment-on-AIGC?style=flat-square&labelColor=343b41"/></td>
 </tr>
 </tbody>
</table>

Long-Tail-image-aesthetics-and-quality-assessment

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README