Comparison of Network Architectures for Semantic Segmentation

Abstract Semantic segmentation is an important task in computer vision applications like autonomous driving, remote sensing, and medical image processing. Recently deep convolutional neural networks have become a standard in semantic segmentation tasks. The encoder-decoder architecture, an example of which is the UNet, has become well established and widely used. However, a two stream architecture working at different resolutions, an example of which is the DDRNet, has shown promising results. This paper aims to compare the performance of encoder-decoder architecture to the two-stream architecture by training a UNet and a DDRNet under the same conditions over different datasets. The results of this paper showed that in all cases, with and without pre-training, over all datasets the two-stream architecture outperformed the encoder-decoder architecture. Although well established, the UNet, seems to have inferior performance when compared to the newer DDRNet.