tf-bilinear-cnn

Requirements

1. TensorFlow with gpu support, my TensorFlow version is 1.80

2. You shold have at least one dataset to run the training FGVC-Aircraft, Caltech-UCSD Birds-200-2011, Cars Dataset

3. Download Pretrained VGG Model

Build the Dataset

1. Change the dataset path in data/aircraft_data.py, data/cub200_data.py and data/standford_cars.py

$ python data/build_aircraft_data.py 

$ python data/build_cub200_data.py

$ python data/build_standford_cars.py

Training & Testing

1. Change the tfrecord path in data/dataset_factory.py

Command for training

$ python train.py

Command for testing

$ python test.py

My Results

Dataset	CUB200	FGVC-Aircraft	Standford Cars
Accuracy	82.6%	84.2%	88.5%

Something interesting!

After training is finished, I visualize some activation maps after vgg/pool5 layer:

References

@inproceedings{lin2015bilinear,
    Author = {Tsung-Yu Lin, Aruni RoyChowdhury, and Subhransu Maji},
    Title = {Bilinear CNNs for Fine-grained Visual Recognition},
    Booktitle = {International Conference on Computer Vision (ICCV)},
    Year = {2015}
}

I also steal some ideas from https://github.com/HaoMood/bilinear-cnn and https://github.com/abhaydoke09/Bilinear-CNN-TensorFlow

Recent Change

I refactor the training/testing code with tf.estimator API