MarUCOD: Unknown but Concerned Object Detection in Maritime Environments

Detection of unknown but concerned objects such as diverse unknown suspicious objects in the sea area is a critical problem in military defense applications, but the problem is challenging because 1) a pre-trained detector cannot easily detect unknown but concerned objects, 2) it detects too many unconcerned objects such as usual objects in the coastal land area, and 3) we cannot easily establish a well-performing discriminator to divide the detected objects into unknown and known objects, as well as concerned and unconcerned objects because the unknown objects are not available, whereas the concerned and unconcerned objects are not clearly defined. To tackle this challenge, this paper proposes a real-time framework for unknown but concerned object detection in maritime environments by integrating object detection, segmentation, and out-of-distribution (OOD) detection techniques. In our framework, an object detector finds all object-like foregrounds by setting a low threshold and a segmentation deep-learning network filters out unconcerned foregrounds detected in the coastal land area. After that, to discriminate known or unknown objects among concerned objects detected in the sea area, a discriminator performs unsupervised OOD detection using bisecting K-means clustering. To boost the performance of the proposed framework, we apply a pre-processing scheme and a contrastive separation loss for segmentation. The proposed framework achieves a high detection rate of unknown but concerned objects with minimal detection of unconcerned objects (i.e., minimal false positives), surpassing baseline methods and demonstrating potential for enhanced maritime safety and security. The codes are open at https://github.com/AIX-Coast-Defense-PIL/MarUCOD.