Remote SAMsing Enhances SAM2 for Remote Sensing Segmentation

Remote SAMsing, an innovative open-source pipeline, tackles two major challenges of Meta's Segment Anything Model 2 (SAM2) in the context of extensive remote sensing images. The first challenge involves a trade-off between quality and coverage in SAM2's mask generator; strict thresholds yield accurate masks but leave much of the image unsegmented, while looser thresholds enhance coverage at the expense of mask quality. The second challenge arises from the necessity to tile large images, which can fragment objects at tile edges. Remote SAMsing addresses both issues without altering SAM2 or needing training data. It employs a multi-pass algorithm that processes each tile multiple times, applying black paint to accepted masks to simplify the scene and adjusting quality thresholds only when coverage improvements plateau. To ensure spatial consistency, it utilizes contextual padding and a parameter-free best-match merge to reconstruct objects across tile edges. This pipeline is intended to be open-source for remote sensing use.