Local-to-Global Point Supervised Object Detector via Aggregation of Discriminative Parts

Advanced fully supervised detectors benefit from abundant bounding-box annotations which accurately cover multi-scale objects. However, for point supervised object detectors (PSOD), each object is annotated by a single point without the information of scales. Some scholars have started to represent the scale of the objects through point-to-box regression, yet the accuracy is constrained by manual heuristic algorithms or local part activations. In this paper, we propose a Local-to-global Point Supervised Object Detector called LPSNet, which can adaptively generate globally aware pseudo bounding boxes. Initially, Point-level Prediction (PLP) precisely identifies the object’s location at a point level. Subsequently, Box-level Prediction with Aggregation of Discriminative Parts (BPAP) dynamically performs regression from points to component-level proposals, and consolidate those part proposals into global ones. Finally, under the supervision of the pseudo proposals, LPSNet with region proposal network and detection head attached obtains detection results. Extensive experiments on the MSCOCO datasets underlines the superiority of our approach, outperforming other available PSOD methodologies.