Treffer: Region-aware unknown information enhancement for open world object detection.

Open World Object Detection (OWOD) focuses on detecting unknown objects of images on the basis of guaranteeing the detection of known objects. However, there still exist the weak transfer ability from the known to the unknown objects and low recognition performance between backgrounds and unknown objects in OWOD. Aiming at the issues, we propose a novel open world object detector with well region-aware and unknown-enhanced properties. In the detector, we design a region-aware redundant box filter by filtering overlap boxes of the known objects and redundant boxes of the unknown objects, and the filter can alleviate the object incompletion and the object intensiveness within proposals. Besides, the detector further constructs a unknown class information enhancement module, which can achieve the stable inference of proposal confidence. For the module, the completeness of the unknown objects is determined by integration confidence, and the unknown objects with the high completeness can be embedded into ground truth boxes. The embedded ground truth boxes are further returned to the objective header, which can further enhance the information of the unknown objects and can reduce the false detection rate of the unknown objects. Extensive experiments on COCO-OOD and COCO-Mix datasets show that the detector is superior to some state-of-the-art detector of OWOD. [ABSTRACT FROM AUTHOR]