2anchao / VovJpu

I use the vovnet27 as the backbone to extract features, at the same time,Jpu is used to get more semantic information that combine with vovnet27.

VovJpu

Model

I use the vovnet27 as the backbone to extract features.

-->>The vovnet27 is described in paper:An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection.the download link is :https://arxiv.org/abs/1904.09730

Jpu is used to get more semantic information that combine with vovnet27.

-->>The Jpu is described in paper:FastFCN: Rethinking Dilated Convolution in the Backbone for Semantic Segmentation.the download link is :http://export.arxiv.org/abs/1903.11816

Model structure

OSA: is the core structure for vovnet.

JPU: is the core structure for FastFCN.

Environment

If you want to implement this project.the environment need build as follow:

python==3.6

torch==1.1.0

numpy

matplotlib

tensorboardX

Script interpret

The script dataprocess.py is for data read,it's actually a iterable.

The script metrics.py is defined miou.

The script vov_jpu.py is Vovnet27 combine the Jpu.

The script train.py is for train the model.

Train

I trained 120 epochs.bitch size is 8. when you establish the environment,then can implement this project in terminal by "python train.py"

Attention

The project was completed by me independently for academic exchange. For commercial use, please contact me by email [e-mail].