Single-Shot Refinement Neural Network for Object Detection

Abstract

This is a MxNet_gluon re-implementation of Single-Shot Refinement Neural Network for Object Detection.

This project is based on MxNet_GluonCV, and completed by YangJirui.

Train on VOC 2007+2012 trainval and test on VOC 2007 test.

Comparison

use voc2007_metric

SSD 300(vgg-16): 77.2 mAP
Paper's refineDet 320(vgg-16): 80.0 mAP
Our refineDet 320(vgg-16) : 78.9 mAP

Our re-implementation of RefineDet performs about 1mAP lower than Paper's.
We will try our best to fix it in the future.

My Development Environment

1、python2.7 (anaconda recommend)
2、cuda9.0 (cuda 8.0 may cause Nan when training)
3、opencv(cv2)

Train

1、train

cd $PATH_ROOT/scripts/detection/refineDet
./train_scipt.sh