Lightweight learned point-line matcher for viewpoint-robust line matching

Develop a lightweight learned point-line matcher to improve the robustness of UPAL's line matching, particularly under viewpoint changes where matching based solely on descriptors at line endpoints is unreliable.

Background

UPAL does not use a dedicated line descriptor; instead, it matches line segments using the descriptors extracted at their two endpoints. The supplementary qualitative analysis reports that this strategy works under illumination changes in structured man-made scenes but fails substantially under viewpoint changes because line-segment endpoints produced by LSD are not necessarily repeatable and can shift along the segment.

The authors further identify the simple mutual-nearest-neighbor matcher as a limitation and explicitly leave the development of a lightweight learned point-line matcher for future work. Such a matcher is intended to improve the robustness of point-line matching, especially for viewpoint-varying image pairs.

References

The simple mutual-nearest-neighbor matcher used here further limits robustness, and we expect a lightweight learned point-line matcher, which we leave to future work (\cref{sec:conclusion}), to improve these results.

— Unified and Efficient Point-Line Local Features  (2608.19894 - Costa et al., 20 Aug 2026) in Supplementary Material, Section “Point and line matching qualitative examples” (see also Section Conclusion)