Real-Time Loop Closure Detection in Visual SLAM via NetVLAD and Faiss

TL;DR

The paper tackles loop closure detection (LCD) in SLAM under appearance changes by evaluating NetVLAD as an LCD descriptor and addressing real-time constraints with Faiss-accelerated retrieval. It introduces a Fine-Grained Top-K precision-recall framework tailored for LCD’s multi-match setting and compares NetVLAD to traditional DBoW on the KITTI dataset. Results show NetVLAD delivers higher accuracy and robustness with competitive real-time performance, making it a practical drop-in LCD module. The work thus bridges Visual Place Recognition advances and LCD requirements, offering a scalable approach for robust SLAM in dynamic environments.

Abstract

Loop closure detection (LCD) is a core component of simultaneous localization and mapping (SLAM): it identifies revisited places and enables pose-graph constraints that correct accumulated drift. Classic bag-of-words approaches such as DBoW are efficient but often degrade under appearance change and perceptual aliasing. In parallel, deep learning-based visual place recognition (VPR) descriptors (e.g., NetVLAD and Transformer-based models) offer stronger robustness, but their computational cost is often viewed as a barrier to real-time SLAM. In this paper, we empirically evaluate NetVLAD as an LCD module and compare it against DBoW on the KITTI dataset. We introduce a Fine-Grained Top-K precision-recall curve that better reflects LCD settings where a query may have zero or multiple valid matches. With Faiss-accelerated nearestneighbor search, NetVLAD achieves real-time query speed while improving accuracy and robustness over DBoW, making it a practical drop-in alternative for LCD in SLAM.

Figures (11)

• Figure 1: Monocular SLAM estimated trajectory for KITTI sequence 00 without Loop Closure Detection. Due to the lack of loop closure constraints, the estimated trajectory suffers from accumulated drift, leading to localization errors. Since this is a monocular SLAM system, the trajectory is only up to an unknown scale and the absolute meter values on the axes are not accurate.
• Figure 2: Monocular SLAM estimated trajectory and ground truth trajectory for KITTI sequence 00 with Loop Closure Detection. Since this is a monocular SLAM system, the estimated trajectory is only up to an unknown scale, and is scaled to best match the ground truth.
• Figure 3: Similarity heatmap for KITTI sequence 00 based on NetVLAD Euclidean Distance. Darker regions indicate smaller distances (higher similarity) between frames. The diagonal represents self-to-self and adjacent frame matches, which are expected to have high similarity and should be ignored for loop closure detection.
• Figure 4: Pose ground truth clustering result for KITTI sequence 00. Frames with small relative pose differences form natural clusters.
• Figure 5: SLAM pipeline with NetVLAD-based Loop Closure Detection. The NetVLAD-based a loop-closure detection module replaces traditional DBoW-based methods, providing a more robust and accurate alternative while maintaining real-time feasibility through Faiss-based retrieval.