CLCT: Complementary Local Consensus Transformer for Two-View Correspondence Pruning

Correspondence pruning extracts accurate correspondences (inliers) from initial ones, playing a crucial role in computer vision and robotics. This process is challenging due to the dominance of outliers and uneven correspondence distributions, which prevent fixed neighbor selection from reliably capturing local consistency and hinder the effective integration of local consistency into global modeling. To tackle these challenges, we introduce local consistency into the Transformer as a guidance signal for the first time, which allows it to adaptively perceive and enforce local matching constraints and consequently enhances global relationship modeling under uneven correspondence distributions. Motivated by this, we propose the Complementary Local Consensus Transformer (CLCT), which captures complementary local consistency by the Adaptive Feature Pruning (AFP) Graph and the Dynamic Weighted (DW) Graph and integrates them into Transformers to enhance global modeling. Specifically, the AFP graph adopts an adaptive pruning strategy to select high-quality correspondences with dynamic neighborhood adjustment, while the DW graph employs a Feature-Aware Distance block with Structural Salient Convolution to refine weighted neighborhood relationships. These graphs guide AFPTransformer and DWGTransformer, enabling a cooperative “point selection + edge weighting” mechanism that integrates local reliability with global reasoning. Experimental results demonstrate that the proposed CLCT outperforms existing methods across multiple tasks.