Web-Supervised Network for Fine-Grained Visual Classification

Introduction

This is the source code for our paper Web-Supervised Network for Fine-Grained Visual Classification

Network Architecture

The architecture of our proposed model is as follows

Installation

After creating a virtual environment of python 3.7, run pip install -r requirements.txt to install all dependencies

How to use

The code is currently tested only on GPU

• Data Preparation

  • Download data into project root directory and uncompress them using

    wget https://wsnfg-sh.oss-cn-shanghai.aliyuncs.com/web-bird.tar.gz
    wget https://wsnfg-sh.oss-cn-shanghai.aliyuncs.com/web-car.tar.gz
    wget https://wsnfg-sh.oss-cn-shanghai.aliyuncs.com/web-aircraft.tar.gz
    tar -xvf web-bird.tar.gz
    tar -xvf web-car.tar.gz
    tar -xvf aircraft-car.tar.gz
    

• Demo

  • If you want to quickly test the performance, please follow subsequent steps

    • Download model using

     wget https://smoothlearning.oss-cn-hongkong.aliyuncs.com/resnet50_81.pth
    

    • Choose a dataset, create soft link to dataset by

    ln -s web-bird bird
    

    • Modify CUDA_VISIBLE_DEVICES to proper cuda device id in run_demo.sh

    • Activate virtual environment(e.g. conda) and then run the script

    bash run_demo.sh
    

• Source Code

  • If you want to train the whole network from begining using source code on the web fine-grained dataset, please follow subsequent steps

    • Choose a dataset, create soft link to dataset by

    ln -s web-bird bird
    ln -s web-car car
    ln -s web-aircraft aircraft
    

    • Modify CUDA_VISIBLE_DEVICES to proper cuda device id in run_train_resnet.sh

    • Modify data to the dataset you want to use in run_train_resnet.sh

    • Activate virtual environment(e.g. conda) and then run the script

    bash run_train_resnet.sh
    

  • We recommend you use Resnet model, because it's easier to train. Resnet-18 shows similar performance with BCNN but needs much less time to train. Resnet-50 has better performance. If you want to train the BCNN model, use run_train_bcnn.sh.

  

Citation

If you find this useful in your research, please consider citing:

@inproceedings{zhang2020web,
title={Web-Supervised Network for Fine-Grained Visual Classification},
author={Chuanyi Zhang, Yazhou Yao, Jiachao Zhang, Jiaxin Chen, Pu Huang, Jian Zhang, Zhenmin Tang},
booktitle={IEEE International Conference on Multimedia and Expo (ICME)},
year={2020}
}