Enhancing Semi-Dense Feature Matching Through Probabilistic Modeling of Cascaded Supervision and Consistency

Hongchang Min, Yihong Tang, Qiankun Li, Zengfu Wang

Abstract: Local feature matching, which identifies correspondences between image pairs, remains a fundamental challenge in computer vision. Current methods usually utilize multi-scale feature fusion to refine reference areas and filter out irrelevant features. However, relying solely on agent loss for supervising upper-level features can reduce refinement accuracy. In addition, the variance in significance among features within the reference region is often overlooked. In this paper, we propose an approach termed Cascaded Supervision-Neighborhood Consistency-Probabilistic Modeling that generates more accurate reference ranges for feature matching. Specifically, the proposed method first directs cascading supervision of the matching results at various scales, enabling more precise refinement of regions. Then, it aggregates matching results at each scale to maintain neighborhood consistency. Finally, probabilistic modeling of the refined reference region is employed, focusing more on relevant features. Extensive experiments conducted on four popular benchmarks demonstrate that our method achieves state-of-the-art and comparable performance.