Global Pooling, More than Meets the Eye: Position Information is Encoded Channel-Wise in CNNs, ICCV 2021

Global Pooling, More than Meets the Eye: Position Information is Encoded Channel-Wise in CNNs
Md Amirul Islam*, Matthew Kowal*, Sen Jia, Konstantinos G. Derpanis, Neil Bruce

Channel-wise Position Encoding

1. Train and Test GAPNet for location classification or image recognition using the following commands:

        cd channel-wise-position-encoding/
        python trainval_gapnet.py 
        python test_gapnet.py 
   

2. Train and Test PermuteNet for location classification or image recognition using the following commands:

        cd channel-wise-position-encoding/
        python trainval_permutenet.py 
        python test_permutenet.py 
   

Learning Translation Invariant Representation

Code coming soon!

Targeting Position-Encoding Channels

Identify and Rank the position encoding channels followed by targeting the ranked channels using the following commands:

        cd position_attack/
        bash run_rank_target_neurons.sh

Please download the DeepLabv3-ResNet50 model trained on Cityscapes from Dropbox and put it under ./position_attack/checkpoints/

Download the cityscapes dataset and change the dataset root path accordingly!

BibTeX

If you find this repository useful, please consider giving a star ⭐ and citation 🦖

  @InProceedings{islam2021global,
   title={Global Pooling, More than Meets the Eye: Position Information is Encoded Channel-Wise in CNNs},
   author={Islam, Md Amirul and Kowal, Matthew and Jia, Sen and Derpanis, Konstantinos G and Bruce, Neil},
   booktitle={International Conference on Computer Vision},
   year={2021}
 }