Learning arbitrary-shape object detector from bounding-box annotation by searching region-graph

Arbitrary-shape is argued more precise than bounding-box for object detection. However, an arbitrary-shape detector usually requires pixel-level human annotation, which is very expensive and hardly afforded for any real-world application. On the other hand, bounding-box is much easier than pixel-wise segmentation in human labeling. In this paper we aim to realize the arbitrary-shape detection from bounding-box human annotation. To this end, we propose location positiveness, which encodes the information of bounding-box annotation to help obtain region annotation. In addition, we propose two graph-based methods to embed the location positiveness, which enable more accurate model trained from simpler annotation. Experimental results validate the performance of our method.