/BCNet

Code for "Low Shot Box Correction for Weakly Supervised Object Detection"

Primary LanguagePythonMIT LicenseMIT

Low Shot Box Correction for Weakly Supervised Object Detection

This code repo is built on faster-rcnn.pytorch.

The final camera-ready paper is now available at IJCAI proceeding

Installation and Preparation

Firstly, clone the code

git clone https://github.com/ptx9363/BCNet.git

and then follow faster-rcnn.pytorch 's preparation to install the environment and dependency. This repo's specific dependencies are shown below:

  • Python 3.5.6
  • Torch 0.4.1
  • Torchvision 0.2.1
  • Numpy 1.15.4

Dataset

We use VOC2007 dataset in our most experiments. We have run weakly-supervised method, OICR, to provide pseudo bounding boxes for images in VOC2007. Some of our experiments are trained from weakly pre-trained models. In general, we provide all of pretrained models and generated labels here.

  • VOC2007 dataset with pseudo labels, data
  • Pretrained models, models
  • Edge boxes proposals, data

the final data folder should be placed like:

BCNet/data/pretrained_model
      data/VOCdevkit/VOC2007
      data/edge_boxes_data

Train and Test

Before training, the cuda libs are required to compiled by:

pip install cython cffi

cd libs

./setup.sh

From now, we have provided train&test code for BCNet with multi-stage and image-level regularization. Just run:

./train_test_vgg16.sh

All of the model modules are avaiable now while more train&test scripts will be released soon.

Citation

@article{jjfaster2rcnn,
    Author = {Jianwei Yang and Jiasen Lu and Dhruv Batra and Devi Parikh},
    Title = {A Faster Pytorch Implementation of Faster R-CNN},
    Journal = {https://github.com/jwyang/faster-rcnn.pytorch},
    Year = {2017}
}

@inproceedings{renNIPS15fasterrcnn,
    Author = {Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun},
    Title = {Faster {R-CNN}: Towards Real-Time Object Detection
             with Region Proposal Networks},
    Booktitle = {Advances in Neural Information Processing Systems ({NIPS})},
    Year = {2015}
}