Computer Vision News - June 2018

Every month, Computer Vision News reviews a research paper from our field. This month we have chosen to review a new improved version of Polygon-RNN (presented at CVPR 2017 ): Efficient Annotation of Segmentation Datasets with Polygon-RNN++ . We are indebted to Sanja Fidler and her team ( David Acuna , Huan Ling and Amlan Kar ), for allowing us to use images from the paper to illustrate this review. Their article is here and their code is here . This work has been selected for CVPR 2018 , where it will presented during the poster session of Tuesday at 10:10am. Introduction: Manual labeling of images with object masks is tedious and almost prohibitively time-consuming. The authors propose their Polygon-RNN++ , a new improved version of Polygon-RNN (presented at CVPR-2017), to help produce polygonal annotations of objects. Polygon-RNN++ produces a polygon by serially iterating vertex by vertex . This serial vertex prediction method of the model allows easily incorporating a human in the loop. The user can propose a vertex correction, which is then fed back into the model, which will use it to re-predict more correctly all other vertices. The model includes several important improvements: 1) new CNN encoder architecture, 2) Reinforcement Learning paradigm, and 3) use of a Graph Neural Network for high resolution results. The authors evaluated Polygon-RNN++ on the Cityscapes dataset and showed that it outperforms the Polygon-RNN, both for automatic evaluation (10% improvement) and interactive evaluation (requiring 50% fewer clicks by annotators). Moreover, Polygon- RNN++ shows powerful generalization capabilities and improvements over existing pixel-wise methods. 4 Research: Polygon-RNN++ Research by Assaf Spanier Computer Vision News “A model for object instance segmentation that can be used to interactively annotate segmentation datasets.”