search

zhreshold / mxnet-ssd

MXNet port of SSD: Single Shot MultiBox Object Detector. Reimplementation of https://github.com/weiliu89/caffe/tree/ssd
MIT License
764 stars 337 forks source link

how to futher improve mAP with training from scratch? #116

Open GarrickLin opened 7 years ago

GarrickLin commented 7 years ago

I have trained a mobilenet based ssd from scratch using trainval set in VOC2007/2012. Now it has an mAP at 0.425539 and seems hard to make large improvement. It might be limited without a pretrained model as I also tried to train reduced vgg16 from scratch. Should I pretrain a good init model with a large number of images of dataset such imageNet as a base knowledgement ? Or use a focalloss to futher improve the performance in defeating imbalance ? I just feel confused and any advice will be appreciated !

zhreshold commented 7 years ago

Recently people have successfully trained a densenet based ssd from scratch, which indicates that with properly designed backpropagate routes, we can skip the pretraining step. The paper is called DSOD: Learning Deeply Supervised Object Detectors from Scratch, FYI.

xhzcyc commented 7 years ago

I have read and implemented DSOD, but the result is not good! Have you tried it? @zhreshold