Proposal Generation by Learning Visual Relationship
Yahui Dong, Wenli Zhou · 2018
Proposal generation aims to generate bounding box candidates that contain objects. The key for proposal generation is to produce very few bounding boxes that can achieve high recall. In this paper, we propose a novel proposal generation algorithm which introduces visual relationship into proposal generator. The intuition is that objects always exist in co-occurrence, which can be clues for generating proposals coherently. Specifically, it is achieved by concatenating a novel relationship network with region proposal network, and jointly training them end-to-end. Experimental results on Visual Relationship demonstrate that our method can achieve better performance compared with state-of-the-art, especially when evaluating at the top limited number of generated proposals.