The Survey Hole Inpainting Problem: A Machine Learning Approach

This work considers the inpainting of missing data in an indoor field, such as geomagnetism and WiFi fingerprints. As opposed to typical image/video inpainting problems, this problem poses several new challenges. First, unlike images with rectangular shapes and fixed RGB channels, indoor geographic data are multi-channeled and highly influenced by building structures. Second, unlike natural objects with fixed shapes, each geographic field is distinct and geographic data are environmentsensitive, following complex physical laws. Consequently, learning from data in other fields is difficult. Third, such data may be obtained from manual surveys and crowdsourcing, which often results in weakly-labeled and noisy datasets. We model our field data as (i) manually surveyed labeled data with holes and (ii) crowdsourced weakly-labeled data without holes. We propose a two-level adversarial regularization inpainting model to conquer these challenges and validate our results with real field data.