Visual place recognition is usually formulated as a general image retrieval problem which suffers from numerous demanding and realistic environment challenges. In this paper, we exploit the particularity of place images which can be surprisingly helpful on place recognition. Specifically, we find that images of identified places can be effectively matched by remarkable regions like building facades under limited geometry and illumination changes. Based on that observation, a novel region mapping based method is proposed to comprehensively tackle the influences caused by geometric and illumination variance as well as irrelevant interference. Given a query image, we extract remarkable regions with color constancy feature performed at processing illumination variant conditions. We leverage a two-fold transformation estimation based verification strategy dealing with geometry transformation caused by viewpoint changes for matching. The experimental results demonstrate that the proposed method is powerful for visual place recognition.
展开▼
