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.

*: 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.

• Chainercv and its dependencies

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

Many implementation is referenced chainercv. Thanks!

• 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