Reviews: Searching for Efficient Multi-Scale Architectures for Dense Image Prediction

In this paper, authors explore the construction of meta-learning algorithms, through defining a viable search space and corresponding proxy tasks that are specifically designed and assessed for the task of dense prediction/semantic segmentation. The proposed method and the resulting network is tested on three different relevant datasets (Cityscapes, PASCAL VOC 12 and PASCAL person part). Main strengths: The method improves the state-of-the-art on three datasets including the competitive cityscapes and PASCAL VOC 2012. Most of the method ingredients (e.g. the proxy tasks) are well-thought-out and explored. The paper is fairly well-written and is easy to follow.