MaskRCNN with Knowledge Distillation

Mask RCNN with pytorch backend for pedestrian detection on Caltech dataset.

Features

• ROI Align layer
• Better default configurations for ResNet18 training
• Support CityPersons

Requirements

1. pytorch

pip install pytorch==0.4.1
pip install torchvision

2. tensorboardX for visualization

pip install tensorboardX==1.2
pip install tensorflow

How to run the code

1. Compiling libs

cd ./libs
make

2. Download Caltech dateset

   Check the Caltech dataset's homepage.

3. Add soft link to the /data folder

cd /data
ln -s /home/data/Caltech caltech

4. Train the teacher model at first

sh exper/caltech/maskrcnn/train_test_resnet50_softmax.sh

5. Then, train the student model

sh /exper/caltech/maskrcnn/train_test_resnet18_softmax.sh

6. You may need to modify the .yaml file for your own training settings.

Reference

@inproceedings{chen2019learning,
    title   =   {Learning Lightweight Pedestrian Detector with Hierarchical Knowledge Distillation},
    author  =   {Chen, Rui and Ai, Haizhou and Shang, Chong and Chen, Long and Zhuang, Zijie},
    booktitle = {IEEE International Conference on Image Processing (ICIP)},
    year    =   {2019}
}