Panoptic segmentation is an important part of developing autonomous surface vehicles, providing detailed information about dynamic obstacles. For the segmentation of the maritime environment, however, the use of semantic segmentation methods without panoptic segmentation prevails. In this thesis, we address this problem and propose a new panoptic segmentation method PSAM based on the agnostic segmentation method SAM. For instance segmentation, we prompt SAM with bounding boxes, predicted by YOLOv7. For semantic segmentation, we develop a new convolutional head in the SAM architecture, where we also add skip connections to the SAM image encoder. PSAM achieves improved results in both panoptic and semantic segmentation. For dynamic obstacles, it achieves 31.3% higher recognition quality RQ than the best panoptic method Mask2Former, and the overall RQ of things and stuff categories is improved by 12.6%. With improved recognition quality, the method achieves a 10.8% higher panoptic quality PQ. In semantic segmentation, PSAM achieves 10% higher precision and 1.5% higher recall for dynamic obstacle regions, compared to the best method KNet. This also improves the F1 score, which is higher by 6%.
|
