Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition

This branch is developed for fine-grained recognition, the related paper is as follows:

Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition[C]
Chaojian Yu, Xinyi Zhao, Qi Zheng, Peng Zhang, Xinge You*
European Conference on Computer Vision. 2018.

Files

• Original Caffe library
• Sum Pooling layer
  • src/caffe/proto/caffe.proto
  • include/caffe/layers/sum_pooling_layer.hpp
  • src/caffe/layers/sum_pooling_layer.cpp
  • src/caffe/layers/sum_pooling_layer.cu
• L2 Normalize layer
  • include/caffe/layers/l2_normalize_layer.hpp
  • src/caffe/layers/l2_normalize_layer.cpp
  • src/caffe/layers/l2_normalize_layer.cu
• Signed Sqrt layer
  • include/caffe/layers/signed_sqrt_layer.hpp
  • src/caffe/layers/signed_sqrt_layer.cpp
  • src/caffe/layers/signed_sqrt_layer.cu
• Example
  • myexample/data/
  • myexample/HBP/ft_last_layer.sh
  • myexample/HBP/ft_last_layer.solver
  • myexample/HBP/ft_last_layer.prototxt
  • myexample/HBP/ft_all.sh
  • myexample/HBP/ft_all.solver
  • myexample/HBP/ft_all.prototxt

Train model

1. The Installation completely the same as Caffe. Please follow the installation instructions. Make sure you have correctly installed before using our code.

2. Download the CUB dataset and VGG16 model for training.

3. Preprocess the CUB dataset and creat list for training set and validation set, place them in myexample/data/

4. Fine tune the last layer

   ./myexample/HBP/ft_last_layer.sh

5. Fine tune the whole network

   ./myexample/HBP/ft_all.sh

Fine tuned model

The fine tuned model is available: google drive baidu skydrive

Citation

If you find our code helpful, please consider citing our work:

@article{yu2018hierarchical,
  title={Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition},
  author={Yu, Chaojian and Zhao, Xinyi and Zheng, Qi and Zhang, Peng and You, Xinge},
  journal={arXiv preprint arXiv:1807.09915},
  year={2018}
}