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

[国内的小伙伴可以看这]This repo contains the official implementation of ELTA of the ICML 2024 paper.

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

Limin Liu*, Shuai He*, Anlong Ming*, Rui Xie, Huadong Ma

Beijing University of Posts and Telecommunications, *Equal contribution

---

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

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