weiaicunzai / pytorch-camvid

This is the Pytorch implementation for Unet

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Pytorch Segmentation

This is a Pytorch implementation for Sementation Networks All experices run on CamVid dataset

Camvid dataset

The original Camvid dataset have 367 training, 233 test images and 101 validation images, total 32 classes. In this project we will use the original training and test images as training set(600 images in total), and the original validation images(101 images) as validation set.We use Camvid dataset provived by fast.ai, they already splited the training and validation dataset for us.

Most image segmentation papers group the similar classes in the original Camvid dataset into one class,resulting total 12 classes:

'Sky', 'Building', 'Pole', 'Road', 'Pavement', 'Tree', 'SignSymbol', 'Fence', 'Car', 'Pedestrian', 'Bicyclist', 'Void'

We will do the same.

Train the Model

GPU

Model was trained on a P100 GPU, you could just simply run

python train.py

TPU

Training on TPU using Pytorch is supported by the Pytorch/xla library, but there is no currently stable release available, the nightly version is highly unstable, so I'll update my TPU training code after the stable release.

Supported Model

unet
segnet

Result

Dataset Network Parameters mIOU
Camvid UNet 34.5M 0.6296
Camvid SegNet 29.4M 0.5913

About

This is the Pytorch implementation for Unet


Languages

Language:Python 100.0%