Weakly supervised learning on road area extraction via classification labels

• Main Author: Liu, Xuanyi
• Other Authors: Long Cheng
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/164103
• Institution: Nanyang Technological University

With the development of remote sensing technology, aerial map applications such as OpenStreetMap have been widely applied in normal life. The lack of pixel-level road area annotations still hinders their further usage. Since annotating roads at pixel level is labor intensive, several data-driven methods have been proposed for this task also, including both unsupervised methods and fully supervised methods. However, the unsupervised methods’ low accuracy and the fully supervised methods’ requirement for a large dataset is concerned. Therefore, weakly supervised road extraction with satisfactory accuracy and fewer supervisions attracts more attention. This thesis presents a novel weakly supervised model named CAM (Class Activation Maps) and Centerline to Road Network (C22RN). C22RN has an encoder and a decoder split into a classification branch and a segmentation branch. The classification branch leverages classification labels such as city labels or road centerline length labels to build CAM. Then a gated fusion mechanism aggregates CAM and the segmentation branch’s pixel-level segmentation results to refine an output, supervised by some pseudo labels generated from road centerlines. Extensive experiments are conducted on CHN6 dataset, which verifies that C22RN outperforms the baseline by 2.4% for IOU.