ChannelDropBlock_Pytorch

Code release for the paper "Channel DropBlock: An Improved Regularization Method for Fine-Grained Visual Classification" (BMVC2021).

Dependencies:

Python 3.6 with all of the pip install -r requirements.txt packages including:

• torch == 1.3.1
• opencv-python

Data

1. Download the FGVC image data. Extract them to data/cars/, data/birds/ and data/airs/, respectively. Organize the structure as follows:

dataset/
    └── train/
         └── class1/
              └── img1.jpg
              └── img2.jpg
              └── ...
         └── ...
     └── test/
         └── class1/
              └── img1.jpg
              └── img2.jpg
              └── ...
         └── ...

Training:

1. For the CUB-200-2011 dataset, run python train_birds+.py --model {resnet50,vgg19} --cdb {none,max_activation,bilinear_pooling} [options: --visualize] to start training.
2. For the Stanford-Cars dataset, run python train_cars.py --model {resnet50,vgg19} --cdb {none,max_activation,bilinear_pooling} [options: --visualize] to start training.
3. For the FGVC-Aircraft dataset, run python train_airs.py --model {resnet50,vgg19} --cdb {none,max_activation,bilinear_pooling} [options: --visualize] to start training.
4. Run python {train_birds+.py, train_cars.py, train_airs.py} --help to see full input arguments.

Visualize:

1. Visualize online attention dropped/remianed feature maps under folder visual/.