Semantic Segmentation Alternative Technique: Segmentation Domain Generation

TL;DR

This paper introduces a novel semantic segmentation method using Generative Adversarial Networks, framing segmentation as a domain transfer problem and employing a feed-forward network to generate masks from images.

Contribution

It proposes an alternative approach to semantic segmentation by utilizing GANs and domain transfer, differing from traditional region-based convolutional methods.

Findings

Demonstrates feasibility of GAN-based segmentation

Achieves promising initial results

Frames segmentation as domain transfer problem

Abstract

Detecting objects of interest in images was always a compelling task to automate. In recent years this task was more and more explored using deep learning techniques, mostly using region-based convolutional networks. In this project we propose an alternative semantic segmentation technique making use of Generative Adversarial Networks. We consider semantic segmentation to be a domain transfer problem. Thus, we train a feed forward network (FFNN) to receive as input a seed real image and generate as output its segmentation mask.