Building High Resolution Maps for Humanitarian Aid and Development with Weakly- and Semi-Supervised Learning

Computer Vision for Global Challenges Workshop at CVPR

By: Derrick Bonafilia, David Yang, James Gill, Saikat Basu

Abstract

Detailed maps help governments and NGOs plan infrastructure development and mobilize relief around the world. Mapping is an open-ended task with a seemingly endless number of potentially useful features to be mapped. In this work, we focus on mapping buildings and roads. We do so with techniques that could easily extend to other features such as land use and land classification. We discuss real-world use cases of our maps by NGOs and humanitarian organizations around the world—from sustainable infrastructure planning to disaster relief. We investigate the pitfalls of existing datasets for building detection and road segmentation and highlight the way that models trained on these datasets—which tend to be highly specific to particular regions—produce worse results in regions of the world not adequately represented in the training set. We explain how we used data from OpenStreetMap (OSM) to train more generalizable models. These models outperform those trained on existing datasets, even in regions in which those models are overfit, and produce these same high-quality results for a diverse range of geographic areas. We utilize a combination of weakly-supervised and semi-supervised learning techniques that allow us to train on the noisy, crowdsourced data in OSM for building detection, which we formulate as a binary classification problem. We then show how weakly supervised learning techniques in conjunction with simple heuristics allowed us to train a semantic segmentation model for road extraction on noisy and never pixel-perfect training data from OSM.