LGT: Light Geometric Transformer for Point Cloud Registration via Token Sampling

Point cloud registration is a fundamental and crucial task in 3D computer vision, playing a pivotal role in various applications such as robotics, autonomous driving, augmented reality, and 3D reconstruction. In recent years, coarse-to-fine registration methods, such as GeoTransformer, have demonstrated significant advancements in precision and robustness by enabling direct point cloud matching without relying on RANSAC. In revisiting the design choices of GeoTransformer, we have introduced several enhancements that improve both efficiency and accuracy. Our analysis revealed substantial spatial redun-dancies among Tokens. To address this, we applied Farthest Point Sampling (FPS) to sparsify Tokens before feeding them into the Transformer, thereby reducing computational overhead. After performing feature pruning within the Transformer, we leveraged an upsampling operation to restore the pruned Superpoint features back to their original resolution. This approach significantly reduces the computational cost of the Transformer while retaining critical geometric information necessary for registration.To validate the effectiveness of our method, we conducted experiments within the GeoTransformer framework. The results demonstrated that our approach achieves a well-balanced trade-off between speed and accuracy. Specifically, while maintaining comparable registration precision, our method achieved a 30% improvement in processing speed, underscoring its strong practical utility.