Semantic Segmentation: A Zoology of Deep Architectures

In this paper we review the evolution of deep architectures for semantic segmentation. The first successful model was fully convolutional network (FCN) published in CVPR 2015. Since then, subject has become very popular and many methods have been published, mainly proposing improvements FCN. We describe detail Pyramid Scene Parsing Network (PSPnet) DeepLabV3, addition to FCN, which provide a multi-scale description increase resolution recent years, reached bottleneck surpassed by transformers from natural language processing (NLP), even though these models are generally larger slower. chosen discuss about Segmentation Transformer (SETR), architecture with transformer backbone. also SegFormer, that includes interpretation tricks decrease size inference time network. networks presented demo come MM-Segmentation library, an open source segmentation toolbox based on PyTorch. propose compare qualitatively individual images, not global metrics databases as is usually case. images outside their training set. invite readers make own comparison derive conclusions. **This MLBriefs article, code reviewed!**