Hawkeye is a unified deep learning based fine-grained image recognition toolbox built on PyTorch, which is designed for researchers and engineers. Currently, Hawkeye contains representative fine-grained recognition methods of different paradigms, including utilizing deep filters, leveraging attention mechanisms, performing high-order feature interactions, designing specific loss functions, recognizing with web data, as well as miscellaneous.


Nov 01, 2022: Our Hawkeye is launched!

Model Zoo

The following methods are placed in model/methods and the corresponding losses are placed in model/loss.

Get Started

We provide a brief tutorial for Hawkeye.


git clone https://github.com/Hawkeye-FineGrained/Hawkeye.git
cd Hawkeye


  • Python 3.8
  • PyTorch 1.11.0 or higher
  • torchvison 0.12.0 or higher
  • numpy
  • yacs
  • tqdm

Preparing Datasets

Eight representative fine-grained recognition benchmark datasets are provided as follows.

FGDataset name Year Meta-class # images # categories Download Link
CUB-200 2011 Birds 11,788 200 https://data.caltech.edu/records/65de6-vp158/files/CUB_200_2011.tgz
Stanford Dog 2011 Dogs 20,580 120 http://vision.stanford.edu/aditya86/ImageNetDogs/images.tar
Stanford Car 2013 Cars 16,185 196 http://ai.stanford.edu/~jkrause/car196/car_ims.tgz
FGVC Aircraft 2013 Aircrafts 10,000 100 https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/archives/fgvc-aircraft-2013b.tar.gz
iNat2018 2018 Plants & Animals 61,333 8,142 https://ml-inat-competition-datasets.s3.amazonaws.com/2018/train_val2018.tar.gz
WebFG-bird 2021 Birds 18,388 200 https://web-fgvc-496-5089-sh.oss-cn-shanghai.aliyuncs.com/web-bird.tar.gz
WebFG-car 2021 Cars 21,448 196 https://web-fgvc-496-5089-sh.oss-cn-shanghai.aliyuncs.com/web-car.tar.gz
WebFG-aircraft 2021 Aircrafts 13,503 100 https://web-fgvc-496-5089-sh.oss-cn-shanghai.aliyuncs.com/web-aircraft.tar.gz

Downloading Datasets

You can download dataset to the data/ directory by conducting the following operations. We here take CUB-200 as an example.

cd Hawkeye/data
wget https://data.caltech.edu/records/65de6-vp158/files/CUB_200_2011.tgz
mkdir bird && tar -xvf CUB_200_2011.tgz -C bird/

We provide the meta-data file of the datasets in metadata/, and the train list and the val list are also provided according to the official splittings of the dataset. There is no need to modify the decompressed directory of the dataset. The following is an example of the directory structure of two datasets.

├── bird
│   ├── CUB_200_2011
│   │   ├── images
│   │   │   ├── 001.Black_footed_Albatross
│   │   │   │   ├── Black_Footed_Albatross_0001_796111.jpg
│   │   │   │   └── ··· 
│   │   │   └── ···
│   │   └── ···
│   └── ···
├── web-car
│   ├── train
│   │   ├── Acura Integra Type R 2001
│   │   │   ├── Acura Integra Type R 2001_00001.jpg
│   │   │   └── ···
│   ├── val
│   │   ├── Acura Integra Type R 2001
│   │   │   ├── 000450.jpg
│   │   │   └── ···
│   │   └── ···
│   └── ···
└── ···

Configuring Datasets

When using different datasets, you need to modify the dataset path in the corresponding config file. meta_dir is the path to the meta-data file which contains train list and val list. root_dir is the path to the image folder in data/. Here are two examples.

Note that the relative path in the meta-data list should match the path of root_dir.

  name: cub
  root_dir: data/bird/CUB_200_2011/images
  meta_dir: metadata/cub

  name: web_car
  root_dir: data/web-car
  meta_dir: metadata/web_car

Note that, for ProtoTree, it was trained on an offline augment dataset, refer to the link if needed. We just provide meta-data for the offline augmented cub-200 in metadata/cub_aug.


For each method in the repo, we provide separate training example files in the Examples/ directory.

  • For example, the command to train an APINet:

    python Examples/APINet.py --config configs/APINet.yaml

    The default parameters of the experiment are shown in configs/APINet.yaml.

Some methods require multi-stage training.

  • For example, when training BCNN, two stages of training are required, cf. its two config files.

    First, the first stage of model training is performed by:

    python Examples/BCNN.py --config configs/BCNN_S1.yaml

    Then, the second stage of training is performed later. You need to modify the weight path of the model (load in BCNN_S2.yaml) to load the model parameters obtained from the first stage of training, such as results/bcnn/bcnn_cub s1/best_model.pth.

    python Examples/BCNN.py --config configs/BCNN_S2.yaml

In addition, specific parameters of each method are also commented in their configs.


This project is released under the MIT license.


If you have any questions about our work, please do not hesitate to contact us by emails.

Xiu-Shen Wei: [email protected]

Jiabei He: [email protected]

Yang Shen: [email protected]


This project is supported by National Key R&D Program of China (2021YFA1001100), National Natural Science Foundation of China under Grant (62272231), Natural Science Foundation of Jiangsu Province of China under Grant (BK20210340), and the Fundamental Research Funds for the Central Universities (No. 30920041111, No. NJ2022028).


View Github