[論文レビュー] Thegra: Graph-based SLAM for Thermal Imagery

Anastasiia Kornilova, Ivan Moskalenko+3 more|arXiv (Cornell University)|Feb 9, 2026

Robotics and Sensor-Based Localization被引用数 0

ひとこと要約

This paper presents a sparse monocular graph-based SLAM system for thermal imagery that leverages general-purpose learned features (SuperPoint detector and LightGlue matcher) with a preprocessing pipeline and a confidence-weighted factor graph to improve robustness and generalization across thermal sensors and environments.

ABSTRACT

Thermal imaging provides a practical sensing modality for visual SLAM in visually degraded environments such as low illumination, smoke, or adverse weather. However, thermal imagery often exhibits low texture, low contrast, and high noise, complicating feature-based SLAM. In this work, we propose a sparse monocular graph-based SLAM system for thermal imagery that leverages general-purpose learned features -- the SuperPoint detector and LightGlue matcher, trained on large-scale visible-spectrum data to improve cross-domain generalization. To adapt these components to thermal data, we introduce a preprocessing pipeline to enhance input suitability and modify core SLAM modules to handle sparse and outlier-prone feature matches. We further incorporate keypoint confidence scores from SuperPoint into a confidence-weighted factor graph to improve estimation robustness. Evaluations on public thermal datasets demonstrate that the proposed system achieves reliable performance without requiring dataset-specific training or fine-tuning a desired feature detector, given the scarcity of quality thermal data. Code will be made available upon publication.

研究の動機と目的

• Address the challenges of thermal imagery for SLAM, including low texture, noise, and NUC artifacts.
• Leverage off-the-shelf, Vis channel-trained features (SuperPoint, LightGlue) with domain-specific preprocessing for thermal data.
• Incorporate keypoint confidence into a weighted factor graph to enhance robustness and tracking stability.
• Demonstrate generalization across diverse thermal sensors and environments without dataset-specific retraining.

提案手法

• Use SuperPoint for keypoint detection and LightGlue for matching in a monocular SLAM pipeline.
• Apply a tailored thermal preprocessing pipeline (CLAHE, histogram equalization, bandpass, edge-preserving smoothing, median filtering) to improve feature detectability.
• Modify map initialization, tracking, and keyframe management to cope with sparse and outlier-prone thermal matches.
• Incorporate a confidence-weighted factor graph optimization using SuperPoint scores to robustify reprojection-based BA.
• Adopt a multi-stage optimization approach including motion-only BA, local BA, and full BA with robust outlier handling.

実験結果

リサーチクエスチョン

• RQ1 How can general-purpose, visible-spectrum-trained detectors and matchers be effectively adapted to thermal imagery without thermal-specific retraining?
• RQ2 What preprocessing and architectural adjustments are necessary to make graph-based monocular SLAM robust on low-texture, noisy thermal data?
• RQ3 Does a confidence-weighted factor graph improve tracking stability and map quality in thermal SLAM across diverse sensors?
• RQ4 How does the proposed system compare to existing thermal and visual SLAM methods in terms of tracking continuity and trajectory accuracy?

主な発見

• The system achieves robust, dataset-general performance on public thermal datasets without dataset-specific training or fine-tuning.
• Preprocessing choices significantly impact feature detection and matching, with a Chambolle-based denoising plus Histogram Equalization and Median filtering yielding the best results.
• The confidence-weighted factor graph improves tracking stability, reducing premature tracking failures compared with uniform weighting.
• Compared to ORB-SLAM3, ROTIO, and DSO on evaluated datasets, the proposed method often completes trajectories with competitive or superior accuracy and greater tracking continuity.
• The approach demonstrates strong generalization across diverse thermal cameras and environments, highlighting the value of combining general-purpose features with domain-aware preprocessing and robust optimization.