Multiple image joint learning for image aesthetic assessment（MILNet）

This repository contains a pytorch implementation of the paper "Multiple Images Joint Learning for Image Aesthetic Assessment"(Subject to ToMM)

We proposed MILNet employ multiple images for aesthetic assessment. First, we adopt semantic information retrieval reference; Then, we use Graph Convolution Network(GCN) to reason the relation of different nodes; Finally, we utilize the improved loss function AdaEMD to stabilize training process.

Pipeline

Requirements

• System: Linux(e.g. Ubuntu/CentOS/Arch), macOS, or Windows Subsystem of Linux (WSL)
• Python version >=3.6
• Pytorch == 1.10.2 Cuda == 10.2
• TensorboardX
• Opencv == 4.5.5

Install

• Clone repo

git clone https://github.com/dafeigediaozhatian/MILNet
cd MILNet

• Install dependencies(pytorch, scikit-learn, opencv-python, pandas. Recommend to use Anaconda.)

# Create a new conda environment
conda create -n MILNet python==3.8
conda activate MILNet

# Install other packages
pip install -r requirements.txt

Dataset

• AVA dataset(19998)

  • Download the original AVA dataset and image featre split into data/. The directory structure should be like:

    data
    	feature_16928_dataloader.py
    	gcn_dataloader_6144.py
    	ILG_train.txt
    	ILG_test.txt
    

• AVA dataset(19928)

  • the same as AVA dataset(19998)

• Extract feature We extract the AVA dataset to MLSP, and feature reduce to 6144 for training. all the files can be download on here,code is #ddkd#.

Training and test

Traning scripts for two datasets can be found in #MILNet.py#. The dataroot argument should be modified to path_to_<dataset_name>. Run the follwing command for training:

# Training and test on AVA_19998
python MILNet.py

Citation

@inproceedings{shi2022multiple,
  title={Multiple image joint learning for image aesthetic assessment},
  author={Tengfei Shi, Zhenyu Wu, Chenglizhao Chen, Aiming Hao, Yuming Fang},
  booktitle={ToMM},
  year={2023}
}