OA-MIL

This repository includes the official implementation of the paper:

Robust Object Detection With Inaccurate Bounding Boxes

European Conference on Computer Vision (ECCV), 2022

Chengxin Liu¹, Kewei Wang¹, Hao Lu¹, Zhiguo Cao¹, and Ziming Zhang²

¹Huazhong University of Science and Technology, China

²Worcester Polytechnic Institute, USA

Paper | Supplementary

Highlights

• Robust: OA-MIL is robust to inaccurate box annotations, and also effective on clean data;
• Generic: Our formulation is general and applicable to both one-stage and two-stage detectors;
• No extra parameters: OA-MIL does not introduce extra model parameters.

Installation

• Set up environment

# env
conda create -n oamil python=3.7
conda activate oamil

# install pytorch
conda install pytorch==1.10.0 torchvision==0.11.0 -c pytorch -c conda-forge

• Install

# clone 
git clone https://github.com/cxliu0/OA-MIL.git
cd OA-MIL

# install dependecies
pip install -r requirements/build.txt

# install mmcv (will take a while to process)
cd mmcv
MMCV_WITH_OPS=1 pip install -e . 

# install OA-MIL
cd ..
pip install -e .

Data Preparation

• Download VOC2007 and COCO datasets. We expect the directory structure to be as follows:

OA-MIL
├── data
│    ├── VOCdevkit
│    │    ├── VOC2007
│    │        ├── Annotations
│    │        ├── ImageSets
│    │        ├── JPEGImages
│    ├── coco
│        ├── train2017
│        ├── val2017
│        ├── annotations
│            ├── instances_train2017.json
│            ├── instances_val2017.json
├── configs
├── mmcv
├── ...

• Generate noisy annotations:

# generate noisy VOC2007 (e.g., 40% noise)
python ./utils/gen_noisy_voc.py --box_noise_level 0.4

# generate noisy COCO (e.g., 40% noise)
python ./utils/gen_noisy_coco.py --box_noise_level 0.4

• Alternatively, the noisy annotation files (coco dataset) we used are available at google drive.

Training

All models of OA-MIL are trained with a total batch size of 16.

• To train OA-MIL on VOC2007, run

sh train_voc07.sh

Please refer to faster_rcnn_r50_fpn_voc_oamil.py for model configuration

• To train OA-MIL on COCO, run

sh train_coco.sh

Please refer to faster_rcnn_r50_fpn_coco_oamil.py for model configuration

Inference

• Modify test.sh

/path/to/model_config: modify it to the path of model config, e.g., ./configs/faster_rcnn/faster_rcnn_r50_fpn_1x_voc_oamil.py

/path/to/model_checkpoint: modify it to the path of model checkpoint

• Run

  sh test.sh

FAQ

• Is OA-MIL applicable to clean data?

  Yes, OA-MIL is applicable to clean data. Here we show some results on the clean VOC2007 and COCO datasets:

  • VOC2007

  Method	mAP@0.5
  FasterRCNN	77.2
  OA-MIL FasterRCNN	78.6

  • COCO

  Method	AP	AP50	AP75
  FasterRCNN	37.9	58.1	40.9
  OA-MIL FasterRCNN	38.1	58.1	41.4

• Where are the noisy annotation files the paper used?

  • The noisy annotation files of the coco dataset are available at google drive;
  • For the GWHD dataset, please refer to this issue.

Citation

If you find this work or code useful for your research, please consider citing:

@inproceedings{liu2022oamil,
  title={Robust Object Detection With Inaccurate Bounding Boxes},
  author={Liu, Chengxin and Wang, Kewei and Lu, Hao and Cao, Zhiguo and Zhang, Ziming},
  booktitle={Proceeding of European Conference on Computer Vision (ECCV)},
  year={2022}
}

Acknowledgement

This repository is based on mmdetection.