WSOD^2: Learning Bottom-up and Top-down Objectness Distillation for Weakly-supervised Object Detection

By Zhaoyang Zeng, Bei Liu, Jianlong Fu, Hongyang Chao, and Lei Zhang

Introduction

This repo is a toolkit for weakly supervised object detection based on mmdetection, including the implementation of WSDDN, OICR and WSOD^2. The implementation is slightly different from the original papers, including but not limited to

• optimizer
• training epoch
• learning rate
• input resolution
• pseudo GTs mining
• loss weight assignment

The baselines in this rpo can easily achieve 48+ mAP on Pascal VOC 2007 dataset. Some hyperparameters are still tuned, they should bring more performance gain.

Architecture

Results

Method	VOC2007 test mAP	VOC2007 trainval CorLoc	VOC2012 test mAP	VOC2012 trainval CorLoc
WSOD2	53.6	71.4	47.2	71.9
WSOD2*	56.0	71.4	52.7	72.2

* denotes training on VOC 07+12 trainval splits

Installation

Please refere to here for installation

Getting Started

1. Download the training, validation and test data, and unzip

mkdir -p $WSOD_ROOT/data/voc
cd $WSOD_ROOT/data/voc
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
tar xf VOCtrainval_06-Nov-2007.tar
tar xf VOCtest_06-Nov-2007.tar

2. Download the ImageNet pre-trained models, selective search boxes and superpixels

bash $WSOD_ROOT/tools/prepare.sh

If you can not access google drive, you also can download the resources from https://pan.baidu.com/s/1htyljhvYz5qwO-4oH8C3wg (password: u5r3) and unzip them, the directory structure should be like

data
  - VOCdevkit
    - VOC2007
      - voc_2007_trainval.pkl
      - voc_2007_test.pkl
      - SuperPixels
    - VOC2012
      - voc_2012_trainval.pkl
      - voc_2012_test.pkl
      - SuperPixels
pretrain
  - vgg16.pth

3. Training a wsod model

bash tools/dist_train.sh $config $num_gpus

4. Evaluate a wsod model

bash tools/dist_test.sh $config $checkpoint $num_gpus --eval mAP

License

WSOD2 is released under the MIT License.

Citing WSOD2

If your find this repo useful in your research, please consider citing:

@inproceedings{zeng2019wsod2,
  title={Wsod2: Learning bottom-up and top-down objectness distillation for weakly-supervised object detection},
  author={Zeng, Zhaoyang and Liu, Bei and Fu, Jianlong and Chao, Hongyang and Zhang, Lei},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  pages={8292--8300},
  year={2019}
}