The Improvement of RoMa Dense Matching Framework With the Introduction of KAN

Dense feature matching aims to establish pixel-level correspondences between image pairs and serves as a fundamental component in tasks such as 3D reconstruction. Although methods like RoMa have achieved remarkable progress, several limitations remain. First, the use of the large-scale DINOv2 model—with approximately 300 million parameters—results in low deployment efficiency and insufficient task-specific optimization for feature matching. Moreover, DINOv2 provides only coarse patch-level features with a fixed spatial resolution of 14×14, which are not well aligned with the dense, high-resolution feature maps produced by CNN-based architectures, thereby introducing a structural bottleneck for precise dense correspondence estimation. Second, current feature pyramid designs treat multi-scale features equally, overlooking the distinct roles of coarse and fine features. To address these issues, this paper introduces two improvements. This work innovatively integrates RKAN-based architectures into a dense feature matching framework and systematically investigates their effectiveness in improving generalization, efficiency, and robustness under limited training data. Benefiting from the curve-based function approximation and favorable inductive bias of RKAN, the proposed RKAN-ResNet backbone achieves strong representational capacity and improved generalization with significantly fewer parameters. Furthermore, independent channel attention is incorporated into each scale of the feature pyramid, enabling the model to adaptively learn the relative importance of different feature channels. Experiments show that RKAN-ResNet achieves comparable matching accuracy to DINOv2 with only 32 million parameters, while improving inference speed by about 30%. After integrating the attention mechanism, all threshold-based metrics further improve, demonstrating a well-balanced contribution between coarse and fine features.