Repulsion Loss based on Faster R-CNN

Introduction

This project is a repulsion loss implementation based on faster RCNN, aimed to recure the thesis "Repulsion loss" CVPR 2018. This project is based on the repo:

• jwyang/faster-rcnn.pytorch, developed based on Pytorch

Process

• change RPN scale to [3,6,9,12,15,18,21,24,27,30,33]

• dilation: remove the fouth maxpooling in vgg16, and add dilation in the next conv

• Ignore handling: lib/model/rpn/lib/model/rpn/anchor_target_layer.py lib/model/rpn/proposal_target_layer_cascade.py

• hard example: lib/datasets/pascal_voc.py change the label; lib/model/rpn/lib/model/rpn/anchor_target_layer.py lib/model/rpn/proposal_target_layer_cascade.py

• reploss: lib/model/faster-rcnn/repulsion_loss.py

Train

python train_vgg_repulsion.py --cuda --mGPUs