StickyPillars: Robust and Efficient Feature Matching on Point Clouds using Graph Neural Networks.

Robust point cloud registration in real-time is an important prerequisite for many mapping and localization algorithms. Traditional methods like ICP and its derivatives tend to fail without good initialization, insufficient overlap or in the presence of dynamic objects. We overcome these drawbacks by introducing StickyPillars, an end-to-end trained 3D feature matching approach based on a graph neural network. We perform context aggregation with the aid of transformer based multi-head self and cross attention. The network output is used as the cost for an optimal transport problem whose solution yields the final matching probabilities. In contrast to state-of-the-art matching methods, our system does not rely on hand crafted feature descriptors or heuristic matching strategies. Our method outperforms state-of-the art matching algorithms like ICP while being suitable for real-time robotics applications. In particular we demonstrate this capability by comparing the translational and rotational error of reconstructed relative poses between two point clouds.