Computer Vision News - October 2018

Every month, Computer Vision News reviews a research paper from our field. This month we have chosen: Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation . We are indebted to the authors of the DeepLabV3+ model ( Liang- Chieh Chen , Yukun Zhu , George Papandreou , Florian Schroff and Hartwig Adam ), for allowing us to use their images to illustrate our review. Their article is here . Semantic segmentation is a subfield of image segmentation, in which the goal is to classify every pixel of the input image with the correct label classification (such as person, dog, cat, etc.), without differentiating separate objects of the same class. That is, for an area of the image with several people in it all pixels will be labelled “person” (there will be no separate label for each person). There are a wide variety of methodologies for solving the semantic segmentation problem (we have reviewed some of them in the past, such as Mask R-CNN in June 2017 ) . The most prevalent dataset used for evaluation of semantic segmentation methods is the PASCAL VOC 2012. Looking at the leaderboard results of PASCAL VOC 2012 it’s clear that for semantic segmentation the series of DeepLab algorithms is dominant. Today we’ll review DeepLabV3+, the fourth version of the DeepLab series, which was presented a few weeks ago at ECCV 2018: as we are writing these lines, it is the top performer on the leaderboard. DeepLabV3+, as mentioned, is a deep learning algorithm for semantic image segmentation with the following features: (1) Atrous convolution allows control of the feature responses’ resolution within different layers in the network. (2) Atrous spatial pyramid pooling (ASPP) to 4 Research: DeepLabV3+ Encoder-Decoder Research by Assaf Spanier Computer Vision News An efficient and better performing encoder-decoder network…