segmentation

Segmentation on Cityscapes

This is a simple and clean implementation of Deeplabv3+ on cityscapes dataset using tensorflow 2.0. It is meant to give a quick run through of the steps required to implement segmentation in TF2.0

Training methodology follows most implementation of random scale cropping and horizontal flipping

Setup

From https://www.cityscapes-dataset.com, download

1. leftImg8bit_trainvaltest
2. leftImg8bit_demoVideo
3. gtFine_trainvaltest

Note that you will need to register.

Training

python train.py --data_dir cityscape_data

cityscape_data points to the root of the downloaded dataset. Training parameters are defined in the file train.py

Saving inference prediction

python eval.py --data_dir cityscape_data

Evaluating with the current training setting could achieve a result similar to the demo