Semantic Segmentation with Extended DeepLabv3 Architecture

In this work, semantic segmentation has been dealt with convolutional neural networks (CNN) which is a widely used recent approach in the field of computer vision. In the experiments using Cityscapes dataset, the images are scaled by various rates and the CNN architecture named DeepLabv3 is trained with different hyperparameters using these images. After the training phase, the success rates of the trained models were compared. The most successful DeepLabv3 model has achieved a success rate of 78.83% on Cityscapes test set. Afterwards, an ensemble of two different DeepLabv3 models and the Extended DeepLabv3 model is tested. In test results, while the success rate remains nearly the same, an increase in classes such as road and sidewalk is observed.