Leveraging Group Annotations in Object Detection Using Graph-Based Pseudo-labeling

Daniel Pototzky, Matthias Kirschner, Lars Schmidt-Thieme

Abstract: We address the problem of dealing with group annotations in object detection where a multitude of items is included in a single bounding box. The standard training protocols in use for most datasets either ignore anchors overlapping with group annotations in the loss function or discard group annotations completely. In this paper, we argue that group annotations contain unexplored potential in many commonly used datasets. We propose to leverage group annotations by generating pseudo-labels on top of them. We develop a novel graph-based pseudo-label generation method that interprets pseudo-label creation as a graph clustering problem that is suited to deal with overlapping objects. In experiments on CityPersons, MS COCO, and a subset of OpenImages we show that our approach outperforms the usual training strategies on the respective datasets when dealing with group annotations as well as other pseudo-label generation methods. We find that the greater the share of group annotations, the larger the increase in performance.