Repeat Fully Convolutional Network (FCN) by Pytoch

Table of Content

• Background
• Requirement
• Usage
• Reference

Background

• Repeat the FCN out of learning and researching purpose, especially for autonomous driving scenario.
• Due to the limitation of my GPU memory (2GB), I only implement a version of FCN with resNet34.
• I train the FCN with CityScape Dataset, the overall mean iou of fine label in CityScape Dataset is about 27%, the benchmark result with resNet101 is 30.4%, thus the performance is acceptable.
• It should be noted that the state-of-the-art performance of mean IoU is 84.5%

Requirement

• matplotlib==3.1.3
• numpy==1.18.1
• torch==1.3.1
• torchvision==0.4.2

Usage

step1: download the dataset

• Download the CityScape Dataset from here
• You will have to create an account with your email address to download the dataset
• Copy the dataset to ./data, the file sturcture of the workspace can be like the following:

step2: train the FCN

• Setting the training parameter (like learning rate, weight decay, number of epoches, batch sizes, etc)

• Just run train.py

step3: evaluate the model

• Just run validation.py

step4: test the model on image from other dataset

• Then run test.py
• Two examples from RobotCar Dataset are provided in ./test

Reference

• Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 3431-3440.