DFL-CNN : a fine-grained classifier

This is a simple pytorch re-implementation of CVPR 2018 Learning a Discriminative Filter Bank Within a CNN for Fine-Grained Recognition.

Introduction:

This work still need to be updated. The features are summarized blow:

• Use VGG16 as base Network.
• Dataset CUB-200-2011, you can split trainset/testset by yourself.Or you can download dataset which has been split directly from BaiduYun Link.
• This work has been trained on 4 Titan V after epoch 120 with batchsize 56, Now I got best result Top1 85.140% Top5 96.237% which is lower than author's. You can download weights from weights.
• Part FCs is replaced by Global Average Pooling to reduce parameters.
• Every some epoches, ten best patches is visualized in vis_result directory, you can put images you want to visualize in vis_img named number.jpg.
• Update: ResNet-101 DFL-CNN and Multi-scale DFL-CNN need to be done.

Algorithms Introduction:

Results and Visualization of ten boxes for discriminative patches:

• This work has been trained on 4 Titan V after epoch 120 with batchsize 56, Now I got best result Top1 85.140% Top5 96.237% which is lower than author's. You can download weights from weights. If use TenCrop transform in code, result can improve further.

• Test Results:

  

• Visualization:

  

Usage:

• Download dataset, you can split trainset/valset by yourself

  wget http://www.vision.caltech.edu/visipedia-data/CUB-200-2011/CUB_200_2011.tgz

• Or you can directly get it from BaiduYun Link
• Then link original dataset to our code root/dataset

  ln -s ./train path/to/code/dataset/train 
  ln -s ./test  path/to/code/dataset/test

• Finally, Train and Test.
• Check you GPU resources and modify your run.sh.

  sh run.sh

Note:

1. Visualization of ten best boxes is saved in vis_result/, img you want to visualize should be put in vis_img/.
2. Weight(checkpoint.pth.tar, model_best.pth.tar) is in weight/.
3. Loss info is saved in log/.