When SAM Meets Sonar Images
Lin Wang, Xiufen Ye, Liqiang Zhu, Weijie Wu, Jianguo Zhang, Huiming Xing, Chao Hu
Abstract
Segment Anything Model (SAM) has revolutionized the way of segmentation due to its remarkable capacity for generalized segmentation. However, SAM’s performance may decline when applied to tasks involving domains that differ from natural images. Nonetheless, by employing fine-tuning techniques, SAM exhibits promising capabilities in specific domains, such as medicine and planetary science. Notably, there is a lack of research on the application of SAM to sonar imaging. In this paper, we aim to address this gap by conducting a comprehensive investigation of SAM’s performance on sonar images. Specifically, we evaluate SAM with various settings on sonar images. Moreover, we fine-tune SAM for sonar images using effective methods both with prompts and for semantic segmentation. The experimental results reveal a substantial enhancement in the performance of the fine-tuned SAM, increasing from 0.24 to 0.75 in mIoU. This underscores the promising potential of SAM for sonar image segmentation applications. Additionally, even when only 2 out of the 11 categories are utilized for training, the model with box prompt sustains an mIoU of 0.69, showcasing its outstanding capability for general segmentation in sonar images. The code is available at https://github.com/wangsssky/SonarSAM.