Home

Awesome

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

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

<p align="left"> <img src="resources/architecture.png" alt="WSOD^2 architecture" width="900px"> </p>

Results

MethodVOC2007 test mAPVOC2007 trainval CorLocVOC2012 test mAPVOC2012 trainval CorLoc
WSOD253.671.447.271.9
WSOD2*56.071.452.772.2

* 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
  1. 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
  1. Training a wsod model
bash tools/dist_train.sh $config $num_gpus
  1. 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}
}