SEM-GAT: Explainable Semantic Pose Estimation using Learned Graph Attention

TL;DR

This paper introduces SEM-GAT, a graph neural network approach that leverages semantic and geometric features for accurate, efficient lidar-based pose estimation with explainability and improved smoothness.

Contribution

The method employs a novel lightweight static graph and cross-graph attention to enhance registration accuracy and interpretability while reducing computational complexity.

Findings

Achieves competitive accuracy on KITTI dataset

Provides higher track smoothness than benchmarks

Uses fewer network parameters

Abstract

This paper proposes a Graph Neural Network(GNN)-based method for exploiting semantics and local geometry to guide the identification of reliable pointcloud registration candidates. Semantic and morphological features of the environment serve as key reference points for registration, enabling accurate lidar-based pose estimation. Our novel lightweight static graph structure informs our attention-based node aggregation network by identifying semantic-instance relationships, acting as an inductive bias to significantly reduce the computational burden of pointcloud registration. By connecting candidate nodes and exploiting cross-graph attention, we identify confidence scores for all potential registration correspondences and estimate the displacement between pointcloud scans. Our pipeline enables introspective analysis of the model's performance by correlating it with the individual contributions of local structures in the environment, providing valuable insights into the system's behaviour. We test our method on the KITTI odometry dataset, achieving competitive accuracy compared to benchmark methods and a higher track smoothness while relying on significantly fewer network parameters.