UPC-Faster-RCNN: A Dynamic Self-Labeling Algorithm for Open-Set Object Detection Based on Unknown Proposal Clustering

Abstract

To promote the development of object detection in a more realistic world, efforts have been made to a new task named open-set object detection. This task aims to increase the model’s ability to recognize unknown classes. In this work, we propose a novel dynamic self-labeling algorithm, named UPC-Faster-RCNN. The wisdom of DBSCAN is applied to build our unknown proposal clustering algorithm, which aims to filter and cluster the unknown objects proposals. An effective dynamic self-labeling algorithm is proposed to generate high-quality pseudo labels from clustered proposals. We evaluate UPC-Faster-RCNN on a composite dataset of PASCAL VOC and COCO. The extensive experiments show that UPC-Faster-RCNN effectively increases the ability upon Faster-RCNN baseline to detect unknown target, while keeping the ability to detect known targets. Specifically, UPC-Faster-RCNN decreases the WI by 23.8%, decreases the A-OSE by 6542, and slightly increase the mAP in known classes by 0.3%.