GOReloc: Graph-based Object-Level Relocalization for Visual SLAM

TL;DR

GOReloc addresses object-level relocalization in visual SLAM by matching frame detections to 3D map objects through semantic uncertainty-aware graphs. It builds source and target graphs, computes node similarities with graph kernels, extracts a search-efficient subgraph, and applies a RANSAC-inspired refinement that weighs residuals by semantic likelihoods and Wasserstein-based distances between dual-quadrics under the estimated pose $\mathbf{T}$. Key contributions include integrating semantic uncertainty and consistency into graph construction and kernel descriptors, a subgraph-extraction strategy that preserves multiple candidates per detection, and a robust pose refinement procedure suitable for real-time operation. Experimental results on TUM RGB-D and ICL-Data Diamond demonstrate improved data association accuracy and pose-success rates, with real-time performance significantly outperforming non-graph baselines and feature-only relocalization. This approach enhances robustness for long-term SLAM in cluttered or dynamic environments and offers a practical pipeline for object-level relocalization in real-world robotics.

Abstract

This article introduces a novel method for object-level relocalization of robotic systems. It determines the pose of a camera sensor by robustly associating the object detections in the current frame with 3D objects in a lightweight object-level map. Object graphs, considering semantic uncertainties, are constructed for both the incoming camera frame and the pre-built map. Objects are represented as graph nodes, and each node employs unique semantic descriptors based on our devised graph kernels. We extract a subgraph from the target map graph by identifying potential object associations for each object detection, then refine these associations and pose estimations using a RANSAC-inspired strategy. Experiments on various datasets demonstrate that our method achieves more accurate data association and significantly increases relocalization success rates compared to baseline methods. The implementation of our method is released at \url{https://github.com/yutongwangBIT/GOReloc}.

Figures (4)

• Figure 1: Our system relocalizes a camera frame in a lightweight, object-level map via graph-based approaches.
• Figure 2: Workflow of our proposed GOReloc. The system relocalizes a camera frame with object detections in an object-level map. After generating the object graphs considering semantic uncertainty and consistency, a subgraph from the target map graph is extracted by selecting association candidates based on node similarity assessment. The final of pose estimation and object associations are refined using a RANSAC-inspired method.
• Figure 3: Illustrations of node similarity assessment via graph kernels and subgraph extraction. The colors of the nodes in the graph distinguish between different categories of detections or objects.
• Figure 4: The qualitative results of object-level data association. In each pair, the left is the frame image, while the right is the map rendering. Correct associations are marked with green lines, and incorrect associations are marked with red lines.