This is a PyTorch implementation of the paper "Three-branch and Mutil-scale learning for Fine-grained Image Recognition (TBMSL-Net)" (Fan Zhang, Guisheng Zhai, Meng Li, Yizhao Liu). Welcome to discuss with us in issues!

• Requirements
• Datasets
• Training TBMSL-Net
• Evaluation
• Model
• Reference

• python 3.7
• pytorch 1.3.1
• numpy 1.17.3
• scikit-image 0.16.2
• Tensorboard 1.15.0
• TensorboardX 2.0
• tqdm 4.41.1
• imageio 2.6.1
• pillow 6.1.0

Download the CUB-200-2011 datasets and copy the contents of the extracted images folder into datasets/CUB 200-2011/images.

Download the FGVC-Aircraft datasets and copy the contents of the extracted data/images folder into datasets/FGVC_Aircraft/data/images)

You can also try other fine-grained datasets.

If you want to train the TBMSL-Net, please download the pretrained model of ResNet-50 and move it to models/pretrained before run python train.py. You may need to change the configurations in config.py if your GPU memory is not enough. The parameter N_list is N1, N2, N3 in the original paper and you can adjust them according to GPU memory. During training, the log file and checkpoint file will be saved in model_path directory.

If you want to test the TBMSL-Net, just run python test.py. You need to specify the model_path in test.py to choose the checkpoint model for testing.

We also provide the checkpoint model trained by ourselves, you can download if from Google Drive for CUB-200-2011 or download from here for FGVC-Aircraft. If you test on our provided model, you will get 89.6% and 94.5% test accuracy, respectively.

If you are interested in our work and want to cite it, please acknowledge the following paper:

@misc{zhang2020threebranch,
    title={Three-branch and Mutil-scale learning for Fine-grained Image Recognition (TBMSL-Net)},
    author={Fan Zhang and Guisheng Zhai and Meng Li and Yizhao Liu},
    year={2020},
    eprint={2003.09150},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}