Decoupled Metric Network for Single-Stage Few-Shot Object Detection

Introduction

Our project is based on the public detection toolbox and benchmark MMDetection v1.1.0.

Quick Start

1. Build

• Clone Code

git clone https://github.com/yrqs/DMNet.git
cd DMNet

• Install MMDetection (see MMDet_README)

2. Prepare Data

• Refer to MPSR. The generated related files can also be downloaded here.
• The final dataset file structure is as follows:

  ...
  configs
  data
    | -- coco
            | -- annotations
                    | -- instances_train2014_base.json
                    | -- instances_valminusminival2014_base.json
                    | -- instances_minival2014.json
                    | -- instances_valminusminival2014.json
                    | -- instances_train2014_*shot_novel_standard.json
                    | -- instances_val2014_*shot_novel_standard.json
            | -- images
                    | -- trainval2014
    | -- VOCdevkit
            | -- VOC2007
                    ...
                    | -- ImageSets
                            | -- Main
                                    | -- trainval_split*_base.txt
                                    | -- trainval_*shot_novel_standard.txt
                                    | -- test.txt
            | -- VOC2012
                    ...
                    | -- ImageSets
                            | -- Main
                                    | -- trainval_split*_base.txt
                                    | -- trainval_*shot_novel_standard.txt
  ...

3. Config files

• Config files are shown below:

  configs
    | -- few_shot
            | -- coco
                    | -- dmnet
                            | -- base.py
                            | -- finetune.py
            | -- voc
                    | -- dmnet_split*
                            | -- base.py
                            | -- finetune.py

4. Training and Finetuning

• Training on base classes:

# remember to change work_dir in dist_train.sh
tools/dist_train.sh config_file num_gpus

• Finetuning on all classes:

# remember to change load_from in config_file
# remember to change work_dir in dist_finetuning.sh
tools/dist_finetuning.sh config_file num_gpus

5. Test

# if test on coco, change '--eval mAP' to '--eval bbox'
tools/dist_test.sh config_file checkpoint_file num_gpus

Acknowledgement

This repo is developed based on MMDetection v1.1.0. Please check them for more details and features.

Citing

If you use this work in your research or wish to refer to the baseline results published here, please use the following BibTeX entries:

@ARTICLE{9721820,
  author={Lu, Yue and Chen, Xingyu and Wu, Zhengxing and Yu, Junzhi},
  journal={IEEE Transactions on Cybernetics}, 
  title={Decoupled Metric Network for Single-Stage Few-Shot Object Detection}, 
  year={2023},
  volume={53},
  number={1},
  pages={514-525},
  doi={10.1109/TCYB.2022.3149825}}