Skip to content

Latest commit

 

History

History
127 lines (87 loc) · 6.11 KB

README.md

File metadata and controls

127 lines (87 loc) · 6.11 KB

Towards Robust Fine-grained Recognition by Maximal Separation of Discriminative Features (ACCV 2020)

Krishna Kanth Nakka and Mathieu Salzmann

Implementation of the ACCV 2020 paper: Towards Robust Fine-grained Recognition by Maximal Separation of Discriminative Features.

Main Paper

Introduction

We introduce an attention-based regularization mechanism that maximally separates the latent features of discriminative regions of different classes while minimizing the contribution of the non-discriminative regions to the final class prediction.

Our approach consists of two modules acting on the features extracted by a backbone network. The attention module extracts attention maps that help the network to focus on the discriminative image regions. The feature regularization module further uses the attention maps to encourage separating the learned prototypes belonging to different classes




Requirements

We provide a docker image shared in this google drive location. Please build this image to get the exact training environment. For the hardware infrastructure, we use NVIDIA Tesla V100 32GB as the GPU resource.

To load the docker image, please use the command below; and push it to the registry

docker load  -i ./path_to_docker/accv.tar

Usage

  1. Install PyTorch

    • The code is tested on Python 3.6, Torch 1.4.0, and NumPy 1.18.1.
    • The code is modified from ProtoPNet.

  2. Clone the repository

    git clone https://github.com/krishnakanthnakka/RobustFineGrained/

  3. Dataset

    • Download the CUB200 dataset. Crop the images using bounding box information.
    • Please put CUB200 test dataset in folder ./datasets/cub200_cropped/test_cropped/ and train dataset in ./datasets/cub200_cropped/train_cropped/. For training, we use same augmentation as ProtoPNet and save in the folder ./datasets/cub200_cropped/train_cropped_augmented/.
    • Please refer to ProtoPNet for preparing the augmented dataset.

  4. Training

    • Normal training
    • python train.py --net=AttProto --mode=normal --backbone=vgg16 --dataset=cub200
    • Adversarial training:
    • python train_robust.py --net=AttProto --mode=robust --backbone=vgg16 --dataset=cub200

  5. Pretrained models

  6. Evaluation of Robust models

    • For testing on robust models with our approach(Ours-A or Ours-FR), please run:

    • # For attacking attention branch that use same architecture as that of AP
python run_adv_attack_AttProto.py -branch=A -dataset=cub200 -config=settings_robust.yaml \
-mode=robust -split=test -backbone=vgg16 -net=AttProto -checkpoint=model.pth -attack=fgsm1_2

# For attacking feature regularization branch
python run_adv_attack_AttProto.py -branch=FR -dataset=cub200 -config=settings_robust.yaml \
-mode=robust -split=test -backbone=vgg16 -net=AttProto -checkpoint=model.pth -attack=fgsm1_2

    • For testing on robust models with Attention Pooling (AP), please run

    •  python run_adv_attack_AP.py -dataset=cub200  -config=settings_robust.yaml \
   -mode=robust -split=test -backbone=vgg16 -net=AP -checkpoint=model.pth -attack=fgsm1_2

    • For testing on robust models with ProtoPNet, please run

    • python run_adv_attack_Proto.py  -dataset=cub200  -config=settings_robust.yaml \
-mode=robust -split=test -backbone=vgg16 -net=Proto -checkpoint=model.pth -attack=fgsm1_2
    • bash run_advexps.sh

      This will produce results in Table 2 in the main paper for VGG16 backbone, which are:

      Model Clean FGSM (1,2) FGSM (1,8) BIM (10,2) BIM (10,8) PGD (10,2) PGD (10,8) MIM (10,2) MIM (10,8)
      AP* 54.9 44.9 24.2 41.9 18.2 41.2 16.9 41.9 18.7
      ProtoPNet* 60.1 44.5 26.9 57.1 10.9 35.9 10.3 37.6 13.5
      Ours-A* 69.3 56.1 34.8 51.7 29.6 50.8 28.0 52.0 32.5
      Ours-FR* 63.0 53.3 37.2 49.4 30.5 48.1 28.4 49.7 31.1

Citation

if you find our work useful in your research, please consider citing:

@InProceedings{Nakka_2020_ACCV,
    author    = {Nakka, Krishna Kanth and Salzmann, Mathieu},
    title     = {Towards Robust Fine-grained Recognition by Maximal Separation of Discriminative Features},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {November},
    year      = {2020}
}

Acknowledgements

We thank ProtoPNet and Attention Pooling authors for the code.