Point-Teaching: Weakly Semi-supervised Object Detection with Point Annotations

Point annotations are considerably more time-efficient than bounding box annotations. However, how to use cheap point boost the performance of semi-supervised object detection is still an open question. In this work, we present Point-Teaching, a weakly- and framework fully utilize Specifically, propose Hungarian-based point-matching method generate pseudo labels for point-annotated images. We further multiple instance learning (MIL) approaches at level images points supervise detector with Finally, simple data augmentation, named Point-Guided Copy-Paste, reduce impact those unmatched points. Experiments demonstrate effectiveness our on few datasets various regimes. particular, Point-Teaching outperforms previous best Group R-CNN by 3.1 AP 5% labeled 2.3 30% MS COCO dataset. believe that proposed can largely lower bar accurate detectors pave way its broader applications. The code available https://github.com/YongtaoGe/Point-Teaching.