Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Revisiting Visual Place Recognition for Joining Submaps in Multimap SLAM

Markus Weißflog, Stefan Schubert, Peter Protzel, Peer Neubert

TL;DR

This paper addresses tracking loss in multimap visual SLAM, which creates disjoint submaps without known relative poses. It introduces a post-processing evaluation pipeline to estimate how modern Visual Place Recognition (VPR) components could improve submap merging without reengineering the SLAM system. Through experiments on the NCLT and Newer College datasets using ORB-SLAM3 with NetVLAD and HDC-DELF, the authors show that modern VPR can enhance map merging, particularly when combined with temporal consistency, though naive descriptor replacement yields limited gains. The work provides a practical method and metrics to gauge VPR potential in SLAM, helping researchers assess benefits before full system integration.

Abstract

Visual SLAM is a key technology for many autonomous systems. However, tracking loss can lead to the creation of disjoint submaps in multimap SLAM systems like ORB-SLAM3. Because of that, these systems employ submap merging strategies. As we show, these strategies are not always successful. In this paper, we investigate the impact of using modern VPR approaches for submap merging in visual SLAM. We argue that classical evaluation metrics are not sufficient to estimate the impact of a modern VPR component on the overall system. We show that naively replacing the VPR component does not leverage its full potential without requiring substantial interference in the original system. Because of that, we present a post-processing pipeline along with a set of metrics that allow us to estimate the impact of modern VPR components. We evaluate our approach on the NCLT and Newer College datasets using ORB-SLAM3 with NetVLAD and HDC-DELF as VPR components. Additionally, we present a simple approach for combining VPR with temporal consistency for map merging. We show that the map merging performance of ORB-SLAM3 can be improved. Building on these results, researchers in VPR can assess the potential of their approaches for SLAM systems.

Towards Revisiting Visual Place Recognition for Joining Submaps in Multimap SLAM

TL;DR

This paper addresses tracking loss in multimap visual SLAM, which creates disjoint submaps without known relative poses. It introduces a post-processing evaluation pipeline to estimate how modern Visual Place Recognition (VPR) components could improve submap merging without reengineering the SLAM system. Through experiments on the NCLT and Newer College datasets using ORB-SLAM3 with NetVLAD and HDC-DELF, the authors show that modern VPR can enhance map merging, particularly when combined with temporal consistency, though naive descriptor replacement yields limited gains. The work provides a practical method and metrics to gauge VPR potential in SLAM, helping researchers assess benefits before full system integration.

Abstract

Visual SLAM is a key technology for many autonomous systems. However, tracking loss can lead to the creation of disjoint submaps in multimap SLAM systems like ORB-SLAM3. Because of that, these systems employ submap merging strategies. As we show, these strategies are not always successful. In this paper, we investigate the impact of using modern VPR approaches for submap merging in visual SLAM. We argue that classical evaluation metrics are not sufficient to estimate the impact of a modern VPR component on the overall system. We show that naively replacing the VPR component does not leverage its full potential without requiring substantial interference in the original system. Because of that, we present a post-processing pipeline along with a set of metrics that allow us to estimate the impact of modern VPR components. We evaluate our approach on the NCLT and Newer College datasets using ORB-SLAM3 with NetVLAD and HDC-DELF as VPR components. Additionally, we present a simple approach for combining VPR with temporal consistency for map merging. We show that the map merging performance of ORB-SLAM3 can be improved. Building on these results, researchers in VPR can assess the potential of their approaches for SLAM systems.
Paper Structure (15 sections, 3 equations, 5 figures, 2 tables)

This paper contains 15 sections, 3 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. ORB-SLAM3
  4. Analysis of the Problem
  5. Loops and Submap Merges in ORB-SLAM3
  6. Performance of Modern Holistic Descriptors
  7. Naive Approach: Only Replacing the VPR Descriptor
  8. Approach
  9. Predicting Submap Adjacency
  10. Evaluation Pipeline
  11. Assumptions and Limitations
  12. Experimental Setup
  13. Results
  14. Discussion
  15. Conclusion

Figures (5)

  • Figure 1: (a) Ground truth trajectory. (b) The output of ORB-SLAM3 is a set of disjoint submaps, with no known transformations between them. (c) The transformations are estimated using VPR (red dotted lines) and temporal consistency (blue dotted lines). (d) The final joined map.
  • Figure 2: Precision-Recall curves for the analyzed descriptors on the NCLT dataset. Better performance is indicated by curves that are near the top right corner of the plots.
  • Figure 3: Sankey diagram of the checks performed by ORB-SLAM3. The width of the arrows without the head corresponds to the number of frames that take this path. Downward pointing arrows mean that the loop/ merge is aborted.
  • Figure 4: The pipeline shows how precision and coverage are calculated using the example of rule VPR. The VPR descriptors, the ORB-SLAM trajectory, and the ground truth are input to the pipeline. $S^F$ and $S$ are continuous matrices, $A$, $A_{gt}$, $R$, and $R_{gt}$ are boolean.
  • Figure 5: Precision-Coverage curves for the analyzed datasets.