Recent work has applied machine learning methods to automatically find and/or assess pedestrian infrastructure in online map imagery (e.g., satellite photos, streetscape panoramas). While promising, these methods have been limited by two interrelated issues: small training sets and the choice of machine learning model. In this paper, aided by the recently released Project Sidewalk dataset of 300,000+ image-based sidewalk accessibility labels, we present the frst examination of deep learning to automatically assess sidewalks in Google Street View (GSV) panoramas. Specifcally, we investigate two application areas: automatically validating crowdsourced labels and automatically labeling sidewalk accessibility issues. For both tasks, we introduce and use a residual neural network (ResNet) modifed to support both image and non-image (contextual) features (e.g., geography). We present an analysis of performance, the effect of our non-image features and training set size, and cross-city generalizability. Our results significantly improve on prior automated methods and, in some cases, meet or exceed human labeling performance.