AFP:Robustizing Object Detection Networks Using Augmented Feature Pooling (ACCV2022, Oral)

This is an official PyTorch impelementation of our paper "Robustizing Object Detection Networks Using Augmented Feature Pooling (ACCV2022, Oral)"

1. Getting Started

Our implementation is based on MMDetection v2.19.0.

To merge code with mmdetection run the following shell.

sh ./merge_mmdet.sh

2. Training

Train Our Augmentation Feature Pooling with Faster RCNN on Our COCO-Rot-Train.

bash ./tools/Faster/dist_faster_Prop_FT4.sh

Train naive Faster RCNN on Our COCO-Rot-Train.

bash ./tools/Faster/dist_faster_Prev.sh

3. Directory Configuration

DET_AFP  
├── mmdet  
├── tools  
├── configs  
├── data  
│   ├── coco  
│   │   ├── annotations  
│   │   ├── train2017  
│   │   ├── val2017  
│   │   ├── test2017  

NOTE: In our implementation coco-rot2 (rotated version of MS COCO) is located in the upper directory as followings:

mmdetection 
├── wrok  
│   ├── DET_AFP  
│
├── data  
│   ├── coco-rot2  
│   │   ├── annotations  
│   │   ├── train2017  
│   │   ├── val2017  
│   │   ├── test2017  

The location of the dataset is arbitrary, but you will need to change the description below according to your location.

data_root = data_root = "../../mmdetection/data/coco_rot2/"

in

configs/_base_/datasets/coco_detection_ROT2.py

4. Pseudo-Code for Our Augmented Feature Pooling

5. Extension to Transformer-Based Backbone

For Swin Transformer, our augmented feature pooling is applied to the feature map immediately after the Swin-Transformer Block of each stage, i.e. just before the Patch Merging.

Citation

If you use this toolbox or benchmark in your research, please cite this project.

@inproceedings{shibata2022robustizing,  
  title={Robustizing Object Detection Networks Using Augmented Feature Pooling},  
  author={Shibata, Takashi and Tanaka, Masayuki and Okutomi, Masatoshi},  
  booktitle={Proceedings of the Asian Conference on Computer Vision},  
  pages={2388--2405},  
  year={2022}  
}