Abstract
Recently, Neural Architecture Search (NAS) has successfully identified neuralnetwork architectures that exceed human designed ones on large-scale imageclassification problems. In this paper, we study NAS for semantic imagesegmentation, an important computer vision task that assigns a semantic labelto every pixel in an image. Existing works often focus on searching therepeatable cell structure, while hand-designing the outer network structurethat controls the spatial resolution changes. This choice simplifies the searchspace, but becomes increasingly problematic for dense image prediction whichexhibits a lot more network level architectural variations. Therefore, wepropose to search the network level structure in addition to the cell levelstructure, which forms a hierarchical architecture search space. We present anetwork level search space that includes many popular designs, and develop aformulation that allows efficient gradient-based architecture search (3 P100GPU days on Cityscapes images). We demonstrate the effectiveness of theproposed method on the challenging Cityscapes, PASCAL VOC 2012, and ADE20Kdatasets. Without any ImageNet pretraining, our architecture searchedspecifically for semantic image segmentation attains state-of-the-artperformance.