P2P-Net

To Run

Create symlink to dataset path

Run ln -s /PATH/TO/ALL/DATASETS/ ./data/
This will create a data folder with symbolic link to the dataset directory that you point to.
Do the same for outputs in /l/users/SOMEWHERE/ to ./output

Copy weights from here . Runs with python 3.8 and pytorch 1.11.0.

Running evaluation

Run python train.py --dataset_name DATASET_NAME --resume PATH/TO/SAVED/WEIGHTS --eval

DATSET_NAMEs are: air, air+car, foodx for FGVC aircrafts, Aircrafts+Stanford Cars combined and Food101 respectively.

Preparation

Benchmarks

CUB_200_2011 (CUB) - http://www.vision.caltech.edu/visipedia/CUB-200-2011.html

Stanford Cars (CAR) - https://ai.stanford.edu/~jkrause/cars/car_dataset.html

FGVC-Aircraft (AIR) - https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/

Unzip benchmarks to "../Data/" (update the variable "data_config" in train.py if necessary).

Training and evaluation

We train the model with 4 V100. The valid batch size is 16*4=64.

python train.py

Performance

Citation

@article{p2pnet2022,
      title={Fine-Grained Object Classification via Self-Supervised Pose Alignment}, 
      author={Xuhui Yang, Yaowei Wang, Ke Chen, Yong Xu, Yonghong Tian},
      journal={arXiv preprint arXiv:2203.15987},
      year={2022},
}

Acknowledgement

This work is supported by the China Postdoctoral Science Foundation (2021M691682), the National Natural Science Foundation of China (61902131, 62072188, U20B2052), the Program for Guangdong Introducing Innovative and Entrepreneurial Teams (2017ZT07X183), and the Project of Peng Cheng Laboratory (PCL2021A07).