Skip to content
Destruction and Construction Learning for Fine-grained Image Recognition
Branch: master
Clone or download
chenyue93
chenyue93 add CL
Latest commit 16b2d15 Apr 3, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
dataset V0.0 Apr 3, 2019
datasets V0.0 Apr 3, 2019
transforms V0.0 Apr 3, 2019
utils add CL Apr 3, 2019
.DS_Store add CL Apr 3, 2019
.gitattributes Initial commit Apr 3, 2019
CUB_test.py V0.0 Apr 3, 2019
LICENSE V0.0 Apr 3, 2019
README.md V0.0 Apr 3, 2019
resnet_swap_2loss_add.py V0.0 Apr 3, 2019
train_rel.py V0.0 Apr 3, 2019

README.md

Destruction and Construction Learning for Fine-grained Image Recognition

By Yue Chen, Yalong Bai, Wei Zhang, Tao Mei

Introduction

This code is relative to the DCL, which is accepted on CVPR 2019.

This DCL code in this repo is written based on Pytorch 0.4.0.

This code has been tested on Ubuntu 16.04.3 LTS with Python 3.6.5 and CUDA 9.0.

Yuo can use this public docker image as the test environment:

docker pull pytorch/pytorch:0.4-cuda9-cudnn7-devel

Citing DCL

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

@article{chen2019dcl,
  title={Destruction and Construction Learning for Fine-grained Image Recognition},
  author={Chen Yue and Bai, Yalong and Zhang Wei and Mei Tao},
  journal={arXiv preprint arXiv:},
  year={2019}
}

Requirements

  1. Pytorch 0.4.0

  2. Numpy, Pillow, Pandas

  3. GPU: P40, etc. (May have bugs on the latest V100 GPU)

Datasets Prepare

  1. Download CUB-200-2011 dataset form Caltech-UCSD Birds-200-2011

  2. Unzip the dataset file under the folder 'datasets'

  3. Run ./datasets/CUB_pre.py to generate annotation files 'train.txt', 'test.txt' and image folder 'all' for CUB-200-2011 dataset

Testing Demo

  1. Download CUB_model.pth from Google Drive.

  2. Run CUB_test.py

Training on CUB-200-2011

  1. Run train.py to train and test the CUB-200-2011 datasets. Wait about half day for training and testing.

  2. Hopefully it would give the evaluation results around ~87.8% acc after running.

Support for other datasets will be updated later

You can’t perform that action at this time.