Comprehensive Attention Self-Distillation for Weakly-Supervised Object Detection

Update(May 30): Please use the latest parallelized version(May30 branch;24G GPU required). The VOC07 is around 57.00% mAP(slightly better than reported and download the model below). For 12G GPU users, please use the master branch(Some functions are cut down to save memory).

This is the official implementation of:

Zeyi Huang', Yang Zou', Vijayakumar Bhagavatula, and Dong Huang, Comprehensive Attention Self-Distillation for Weakly-Supervised Object Detection, NeurIPS 2020, Arxiv version.

Citation:

@article{huang2020comprehensive,
  title={Comprehensive Attention Self-Distillation for Weakly-Supervised Object Detection},
  author={Huang, Zeyi and Zou, Yang and Kumar, BVK and Huang, Dong},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

Installation

Requirements

• Python == 3.7
• Pytorch == 1.1.0
• Torchvision == 0.3.0
• Cuda == 10.0
• cython
• scipy
• sklearn
• opencv
• GPU: TITAN RTX (24G of memory)

Note: To train with GPU of small memory, CASD_IW is partially parallelized. Fully parallelized version is coming soon. Thanks for your patience.

Preparation

1. Clone the repository

   git clone https://github.com/DeLightCMU/CASD.git

2. Compile the CUDA code

   cd CASD/lib
   bash make_cuda.sh

3. Download the training, validation, test data and the VOCdevkit

   mkdir data
   cd data/
   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
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCdevkit_18-May-2011.tar

4. Extract all of these tars into one directory named VOCdevkit

   tar xvf VOCtrainval_06-Nov-2007.tar
   tar xvf VOCtest_06-Nov-2007.tar
   tar xvf VOCdevkit_08-Jun-2007.tar

5. Create symlinks for PASCAL VOC dataset

   cd CASD/data
   ln -s VOCdevkit VOCdevkit2007

6. Download pretrained ImageNet weights from here, and put it in the data/imagenet_weights/

7. Download selective search proposals from here, and put it in the data/selective_search_data/

Training and Testing

Train a vgg16 Network on VOC 2007 trainval

bash experiments/scripts/train_faster_rcnn.sh 0 pascal_voc vgg16

Test a vgg16 Network on VOC 2007 test

bash experiments/scripts/test_faster_rcnn.sh 0 pascal_voc vgg16

Download log and weight from here

Acknowledgement

We borrowed code from MLEM, PCL, and Faster-RCNN.