Benchmarking RefineDet[1] and other SSD (Single shot Detection) Network based on chainer

Including

• DSSD[2]
• SSD with residual prediction module[2]
• ESSD[3]
• RefineDet[1]

Original DSSD is based on ResNet 101. Since memory limitation, only tried on VGG.

Benchmark result

Model name	Base model	Input image size	mAP	mAP(paper)
SSD	VGG16	300x300	77.5	77.5
SSD Plus(Use Residual Prediction module)	VGG16	300x300	78.0	NA
ESSD	VGG16	300x300	78.8	79.4
RefineDet	VGG16	320x320	Now evaluating	80.0

*: I set batchsize to 22 because of memory limitation. The original paper used 32.

*: Some training condition is different from paper.

*: ESSD original paper did 3 stages training (Only SSD, Only extensional module and whole network), but I did whole training only.

*: I may mistook unintensionally.

Requirement

• Chainercv and its dependencies

Usage

git clone https://github.com/fukatani/RefineDet_chainer
cd refinedet
python train.py --model refinedet320 --batchsize 22

Other

Many implementation is referenced chainercv. Thanks!

Reference

• Single-Shot Refinement Neural Network for Object Detection

• DSSD : Deconvolutional Single Shot Detector

• Extend the shallow part of Single Shot MultiBox Detector via Convolutional Neural Network