LightGlueStick: a Fast and Robust Glue for Joint Point-Line Matching

TL;DR

LightGlueStick addresses the need for fast, robust joint point-line matching suitable for real-time and edge deployment. It builds a lightweight transformer-based backbone with a novel Attentional Line Message Passing layer to explicitly encode line connectivity and improve communication between endpoints, while supporting early exit to save computation. The approach achieves state-of-the-art results on ETH3D, HPatches, and ScanNet, and demonstrates strong localization performance with real-time throughput on embedded-like settings. This work enables practical fusion of point and line features for SLAM, visual localization, and other geometry-aware tasks on resource-constrained platforms.

Abstract

Lines and points are complementary local features, whose combination has proven effective for applications such as SLAM and Structure-from-Motion. The backbone of these pipelines are the local feature matchers, establishing correspondences across images. Traditionally, point and line matching have been treated as independent tasks. Recently, GlueStick proposed a GNN-based network that simultaneously operates on points and lines to establish matches. While running a single joint matching reduced the overall computational complexity, the heavy architecture prevented real-time applications or deployment to edge devices. Inspired by recent progress in point matching, we propose LightGlueStick, a lightweight matcher for points and line segments. The key novel component in our architecture is the Attentional Line Message Passing (ALMP), which explicitly exposes the connectivity of the lines to the network, allowing for efficient communication between nodes. In thorough experiments we show that LightGlueStick establishes a new state-of-the-art across different benchmarks. The code is available at https://github.com/aubingazhib/LightGlueStick.

Figures (4)

• Figure 1: LightGlueStick outperforms other state-of-the-art line matchers with a speed comparable to LightGlue lindenberger2023lightglue. The depth adaptivity mechanism allows to predict line matches earlier. The parameter $\alpha$ controls the tradeoff between speed and accuracy. Here, LightGlue lindenberger2023lightglue matches lines using endpoints.
• Figure 2: Overview of the proposed pipeline: LightGlueStick takes as input two images where points and lines have been detected and described. Our Graph Neural Network (GNN) is composed of three stages: 1) Self-Attention, 2) our novel Attentional Line Message Passing, and 3) Cross-Attention. Depending on the achieved matching result at each layer, we can stop the execution to save computation. The last step of Dual-Softmax generates two assignation matrices for points and lines, respectively.
• Figure 3: LightGlueStick successfully predicts matches on rotated images, despite not being trained on them, whereas GlueStick pautrat_suarez_2023_gluestick fails to predict correct line matches.
• Figure 4: Depth Adaptivity of LightGlueStick. LightGlueStick adaptively adjusts its depth based on image difficulty, exiting after the 4th layer for the top pair, the 5th layer for the middle pair, and the 7th layer for the bottom pair. The bottom pair requires more layers due to its smaller visual overlap, making it more challenging to match. Processing times for the top, middle, and bottom pairs are 27ms, 34ms, and 42ms, respectively.