This repository is a toolkit to take a try on semantic segmentation task, we choose Cityscape dataset as example and simplified the traning categories to show how it works.

cd docker
make

It might take few minutes to build up the docker image

To check if the docker image have successful built

docker image | grep aicampus

Assume you have already downloaded the following Cityscape datasets and unzipped them at current directory(path_to/self_driving_car/).

• gtFine_trainvaltest.zip (241MB)
• leftImg8bit_trainvaltest.zip (11GB)

All the data should be arranged in a certain folder in following structure:

 |── self_driving_car
 |   ├── docker
 |   ├── gtFine
 |   ├── leftImg8bit

We choose some of the labels in the original dataset and merge them into 3 main categories to run the test

The original image size of cityscape is 1024 * 2048 , we downsize them to 320 * 640 to accelrate the training process

./import_data.sh

1. Car including car, bus, train, truck, trailer, caravan (no trailer and caravan annation in bdd label annotation set)

2. Person· including person, rider

3. Road·

Double check if there is a file called "class_dict.csv" in downsized output dataset

./train_demo.sh

./predict_demo.sh