Bilinear-CNN

A pytorch implementation of bilinear CNN for fine-grained image recognition in paper Bilinear CNNs for Fine-grained Visual Recognition by Tsung-Yu Lin, Aruni RoyChowdhury, Subhransu Maji.
The framework of the entire model is shown in the figure below:

In this implementation, the base network uses resnet34 structure

Dependencies

python >= 3.5
pytorch >= 0.4
In addition, please add the project folder to PYTHONPATH and pip install the following packages:

• tqdm

Data

Download the CUB_birds images and annotations. Extract them to data/raw/

Training

According to the requirements of this paper, the training of this model is divided into the following two steps:

• The parameters of the pre-training model are fixed, only the last full connection layer is trained. Run the following code.
  python bilinear_ResNet_linear_layer.py
  In this step, the mean value and variance of images data need to be calculated for image preprocessing
• All parameters in the model are trained. Run the following code.
  python bilinear_ResNet_fine_tuning.py

Results

In this implementation, the accuracy of the model on the test dataset of the CUB_bird can reach about 83%. If you want get the model file, you can download it from https://pan.baidu.com/s/1xLsZQCpD-s5FXmyMOGj69w using the extract code 'xxbr', or contact me via email:[email protected]