Endpoint‐Aware Line Segment Matching With Geometry‐Based Prediction

This paper addresses the issue of content inconsistency caused by asymmetry descriptor regions across multiple views when matching the partial correspondent line fragments in man‐made scenes. Existing methods mainly focus on achieving content consistency in common geometric transformation and illumination change by projecting the local information around the interest lines into the descriptor space. However, we find that in practice the asymmetry in local descriptor regions caused by incomplete extraction or occlusion continually occurs and leads to the loss of local information, which cannot be retrieved through projection and destroys the content consistency. Based on this finding, we propose an endpoint‐aware line segment matching method with geometry‐based prediction (GPMatcher) to predict the projected positions of endpoint pairs of line fragments and directly find the most matched line. By constructing content and geometric associations between endpoints and the global image, endpoints that are inexactly regarded as corresponding can be accurately reassigned. Experiment on three representative datasets demonstrates that our method outperforms state‐of‐the‐art in both indoor and outdoor scenarios under a more rigorous evaluation based on downstream tasks.