Gaussian Semantic Field for One-shot LiDAR Global Localization

TL;DR

This work tackles one-shot LiDAR global localization in semantically ambiguous, repetitive environments by introducing Gaussian Semantic Fields (GSF) as an intermediate layer in a 3D scene graph. GSFs are learned with Gaussian processes to model continuous spatial-semantic distributions, enabling fine-grained, metric-aware data association via a 2-Wasserstein similarity between local fields. The pipeline uses a four-stage process—tri-layer scene graphs, GP-based local semantics, substructure matching, and consistency-based inlier unification—to achieve robust 6-DoF localization without extensive map accumulation. Across public datasets, Outram-GSF demonstrates state-of-the-art localization success rates and pose accuracy, particularly excelling in repetitive scenes, and shows a favorable accuracy-runtime trade-off when equipment-intensive refinement is applied.

Abstract

We present a one-shot LiDAR global localization algorithm featuring semantic disambiguation ability based on a lightweight tri-layered scene graph. While landmark semantic registration-based methods have shown promising performance improvements in global localization compared with geometric-only methods, landmarks can be repetitive and misleading for correspondence establishment. We propose to mitigate this problem by modeling semantic distributions with continuous functions learned from a population of Gaussian processes. Compared with discrete semantic labels, the continuous functions capture finer-grained geo-semantic information and also provide more detailed metric information for correspondence establishment. We insert this continuous function as the middle layer between the object layer and the metric-semantic layer, forming a tri-layered 3D scene graph, serving as a light-weight yet performant backend for one-shot localization. We term our global localization pipeline Outram-GSF (Gaussian semantic field) and conduct a wide range of experiments on publicly available data sets, validating the superior performance against the current state-of-the-art.

Figures (4)

• Figure 1: A semantic annotated point cloud of a partially semantically ambiguous scene and two Gaussian semantic fields of the corresponding rectangle areas. The semantic field encodes the metric-semantic distribution of a small area (purple for road, light blue for pole, green for vegetation, and black for traffic sign). Note that the two areas are highly symmetric in geometry, even in instance-level semantics, but naturally differentiable in surrounding vegetation distributions. We leverage this continuous semantic distribution for robust global localization.
• Figure 2: We present Outram-GSF, a one-shot global LiDAR localization pipeline. In stage 1, we generate tri-layer scene graphs for both the current query scan and the target reference map. Around each instance in the object layer $\mathcal{C}_1$, we train sparse Gaussian processes (GPs) using the semantic annotated point clouds in the metric semantic layer. Populations of GPs for metric similarity comparison are generated by linear probing in stage 2. Combined with a semantic stability mask, the populations of GP can be used as a similarity measurement in the Wasserstein sense in stage 3. Given the correspondences established by the above process, inlier correspondences are selected by maximum clique (stage 4).
• Figure 3: Comparison of the semantic reconstruction quality of two downsampling strategies to train the Sparse Gaussian Process in (\ref{['eq:gp_mean_variance']}). We find that our proposed Semantic-based sparsification consistently outperforms random downsampling in terms of the reconstruction mIoU (mean Intersection over Union) under the same sample number condition.
• Figure 4: A visual comparison of the correspondence establishment and global localization performance in a symmetric environment of the proposed method with (right) and without (left) the Gaussian semantic field. Pure instance-level methods result in twisted correspondences (see the reversal correspondences in red ).