SAM-UpBuild: Efficient Automated Building Footprint Update with Limited Labels

Le résumé fourni par la source

Building footprints are pivotal for comprehending urban growth and facilitating sustainable urban development. Given the rapid pace of urban change, it is imperative to maintain up-to-date building records. However, this process is hindered by the time-consuming and labor-intensive task of obtaining labels for new or changed buildings, as well as the high costs associated with training high-precision models. Recently, the Segment Anything Model (SAM) has shown exceptional generalization across diverse scenarios. Building on this, we present SAM-UpBuild, an innovative approach for urban footprint updating. It leverages historical building footprint and integrates the Segmentation Arbitrary Model (SAM) with adaptors for fine-tuning. Our method addresses the challenges of laborious labeling and high training costs by employing a lightweight adaptation strategy that requires only a small number of building labels, e.g. 0.1% of the original training samples. To verify the performance of our proposed method, we conduct experiments on WHU building change detection datasets. Results demonstrate that the proposed method outperforms other state-of-the-art (SOTA) algorithms by a large margin.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.