Semantic-spatial Coupled Segmentation Everything Model for Point-prompted Oriented Object Detection

Remote sensing object detection typically relies on box supervision, which requires great human effort for precise annotation.Recently, the segment anything model (SAM) demonstrated the capability of segmenting objects with simple point prompt learning without fine-tuning.Nevertheless, due to the uncertainty of the single-point prompt, SAM's mask proposals commonly cause ambiguity.In this study, we propose a new approach with the aim of selecting the most suitable mask from proposals according to the point annotation and category.With our approach, the circumscribed rectangle of the estimated pseudo mask can supervise the rotated object detection network training.Further, the semantic-spatial coupled box refinement fuses the classification and spatial information in proposal selection, which brings a more precise proposal.Experiments on the DOTA dataset verify the effectiveness of the proposed approach.