Evaluation of Front-End Feature Matching for Lunar Visual SLAM

Guillem Casadesus Vila; Grace Gao

cs.CVOct 22, 2026

Evaluation of Front-End Feature Matching for Lunar Visual SLAM

Authors: Guillem Casadesus Vila, Grace Gao

Organizations: Department of Aeronautics and Astronautics, Stanford University, Stanford, CA, United States

Abstract

Lunar rovers will primarily rely on cameras for localization until navigation satellite services are deployed, and will continue to need them afterwards for relative navigation, hazard detection, and surface operations. Visual odometry (VO) and Simultaneous Localization and Mapping (SLAM) mostly estimate motion by matching features between frames. However, feature matchers are developed on urban, indoor and driving scenes, planetary studies compare them on orbital registration or within complete SLAM systems, and real lunar imagery with reference poses is scarce, so matching between rover frames on low-texture, harshly lit terrain, and its effect on VO, has not been compared. We benchmark eleven classical and learned detector-matcher combinations on nine synthetic, lunar-testbed, and terrestrial-analogue datasets, and a feature tracker, scoring pairwise pose accuracy against baseline and the trajectory error of the resulting VO. Under a common rotation-only metric, the SuperPoint pairings outperform SIFT with ratio-tested FLANN matching on both lunar testbeds and one analogue site, with smaller or unresolved differences on synthetic data. When rotation and translation direction are scored jointly, this SIFT pairing exceeds both SuperPoint pairings on four synthetic datasets, so synthetic results alone do not establish relative performance on real rover imagery. In stereo VO, the SuperPoint pairings are not consistently more accurate than SIFT, but only they keep tracking on all eight sequential datasets. Pairings with classical detectors depend on lighting while SuperPoint changes little: a low spotlight makes them more accurate than diffuse light on the same testbed views, consistent with more contrast and detected keypoints. SuperPoint with LightGlue is the fastest learned combination (16 ms per frame), and DINOv2 retrieval proposals still need geometric verification.

Figures & tables

Dataset	Type	Conditions	Frames	Reference depth	Ground-truth pose
LuPNT (base)	Synthetic	Sun elevation $15^{\circ}$	4,281	Rendered	Simulator
LuPNT (cam.)	Synthetic	Camera effects	4,276	Rendered	Simulator
LuPNT (high sun)	Synthetic	High sun, elevation $45^{\circ}$	4,281	Rendered	Simulator
LuSNAR (5)	Synthetic	Scene 5	1,599	Rendered	Simulator
LuSNAR (9)	Synthetic	Scene 9	2,087	Rendered	Simulator
PolarTraverse	Lunar testbed	Regolith testbed	11 stations	Laser scan	COLMAP registered to laser scan

TABLE I : Datasets used for evaluation. GNSS: global navigation satellite system.

Fig. 1 : Sample image, reference depth (sky and invalid pixels gray) and top-down trajectory with the sample frame marked, for six of the nine datasets. LunaPolaris shows sequence 07. TLS: terrestrial laser scan.

Fig. 2 : Fraction of mixed-pool pairs per overlap bin (RUBIK [ 60 ] bin edges plus a 0–5 % bin) on six datasets, with median overlap

\tilde{\omega}

and triangulation angle

\bar{\alpha}

		Synthetic					Lunar testbed		Terrestrial analogue
		LuPNT			LuSNAR		PolarTraverse	LunaPolaris	S3LI
Detector	Matcher	(base)	(cam.)	(high sun)	(5)	(9)			(Etna)	(Vulcano)
SIFT	FLANN	58.4	68.5	74.9	67.1	69.6	37.8	31.6	65.0	76.5
SIFT	BruteForce	43.2	53.1	55.9	51.0	53.5	28.8	15.0	47.0	69.3
SIFT	LightGlue	60.4	70.5	75.2	63.2	66.1	44.9	29.2	69.9	78.7
ORB	FLANN	14.4	25.7	26.9	32.8	33.6	3.7	9.9	34.4	49.8

TABLE II : Pose AUC@

5^{\circ}

[%] of eleven combinations and, in italics, four ablation pairings. The pose error is the larger of the rotation and translation-direction errors, except on LunaPolaris and S3LI (rotation alone). Bold/underline = best/second among the eleven. FLANN cells: median of five seeds, spread

\leq 4.9

. Unmarked KLT rows track between the two frames (direct) or through every intermediate frame (chained, “—” without a sequence). AUC@

5^{\circ}

: area under the cumulative pose-error curve up to

5^{\circ}

, FLANN: fast library for approximate nearest neighbors, KLT: Kanade–Lucas–Tomasi, SP: SuperPoint.

		Synthetic					Lunar testbed		Terrestrial analogue
		LuPNT			LuSNAR		PolarTraverse	LunaPolaris	S3LI
Detector	Matcher	(base)	(cam.)	(high sun)	(5)	(9)			(Etna)	(Vulcano)
SIFT	FLANN	79.5	85.2	89.7	92.2	90.2	47.3	31.6	65.0	76.5
SIFT	LightGlue	81.9	86.7	90.3	90.9	88.4	53.2	29.2	69.9	78.7
SP	SuperGlue	89.0	87.2	89.3	90.8	90.3	97.2	56.9	77.0	77.1
SP	LightGlue	88.8	87.6	89.7	91.1	88.4	97.5	55.2	74.8	75.5

TABLE III : Rotation-only Pose AUC@

5^{\circ}

[%] on all nine datasets. Bold/underline = best/second. SIFT+FLANN: median of five seeds. KLT rows, tracked as in Table II , are not marked (“—”: no sequence to chain through). AUC@

5^{\circ}

: area under the cumulative rotation-error curve up to

5^{\circ}

, FLANN: fast library for approximate nearest neighbors, KLT: Kanade–Lucas–Tomasi, SP: SuperPoint.

Fig. 3 : Matches of SIFT+FLANN (left) and SuperPoint+LightGlue (right) on one pair per dataset, from top: LunaPolaris ( $0.34$ m, $4.0^{\circ}$ ), S3LI (Etna) ( $2.78$ m, $29.2^{\circ}$ ) and LuPNT (base) ( $1.05$ m, $0.9^{\circ}$ ). Green = RANSAC inliers, red = outliers. FLANN: fast library for approximate nearest neighbors, RANSAC: random sample consensus.

Fig. 4 : Pose AUC@

5^{\circ}

[%] against the median translation of each frame step [m, log axis], on LuPNT (base), LuSNAR (5), S3LI (Etna) and LunaPolaris. The two real datasets are scored rotation-only. On the synthetic panels the error includes the translation direction, poorly constrained by the small parallax of the shortest steps, so the AUC first rises with the step. At the widest steps, low overlap also fails the rotation. AUC: area under the curve, FLANN: fast library for approximate nearest neighbors, KLT: Kanade–Lucas–Tomasi, SP: SuperPoint.

		Synthetic					Testbed	Analogue
		LuPNT			LuSNAR		LunaPolaris	S3LI
Detector	Matcher	(base)	(cam.)	(high sun)	(5)	(9)		(Etna)	(Vulcano)
SIFT	FLANN	2.3	4.1	5.3	3.7	3.6	0.3	2.8	6.2
SIFT	BruteForce	1.3	1.8	1.9	1.8	2.1	—	0.9	3.7
SIFT	LightGlue	3.0	4.6	5.3	3.2	3.0	0.3	3.8	6.6
ORB	FLANN	—	0.7	1.1	0.9	0.9	—	0.6	1.6

TABLE IV : Operating range

\Delta t^{*}

[m], the largest median translation with Pose AUC@

5^{\circ}\geq 50

%. Pose error as in Table II , rotation alone on LunaPolaris and S3LI. “—”: never. Bold/underline = best/second. FLANN cells: median of five seeds, spread

\leq 0.46

m. AUC: area under the curve, FLANN: fast library for approximate nearest neighbors, SP: SuperPoint.

		Pose AUC@ $5^{\circ}$ [%]			Stereo VO ATE/L [%]
Detector	Matcher	Seq. 01	Seq. 07	Seq. 10	Seq. 01	Seq. 07	Seq. 10
SIFT	FLANN	31.6	30.3	42.1	2 % $\ddagger$	0.6	0.8
SIFT	BruteForce	15.0	20.3	25.9	$<$ 1 % $\ddagger$	11 % $\ddagger$	71 % $\ddagger$
SIFT	LightGlue	29.2	37.4	49.3	2 % $\ddagger$	0.4	1.1
ORB	FLANN	9.9	8.7	14.3	13 % $\ddagger$	9 % $\ddagger$	2.1
ORB	BruteForce	6.7	8.6	14.5	13 % $\ddagger$	9 % $\ddagger$	69 % $\ddagger$

TABLE V : Rotation-only Pose AUC@

5^{\circ}

[%] and stereo VO ATE/L [%] on LunaPolaris sequences 01 (ceiling lights), 07 (low spotlight) and 10 (low spotlight and headlight). Marks as in Table II . Full pools, not pose-matched. FLANN cells: median of five seeds.

\ddagger

Stopped tracking: path coverage. Sequence 01 repeats the LunaPolaris columns of Tables II , III and VI . ATE: absolute trajectory error, ATE/L: ATE as a percentage of path length, AUC: area under the curve, FLANN: fast library for approximate nearest neighbors, SP: SuperPoint, VO: visual odometry.

Fig. 5 : Matches of SIFT+FLANN (left) and SuperPoint+LightGlue (right), with inlier counts and rotation errors, on a $0.34$ m pair from sequence 01 (top) and its pose-matched counterparts in 07 (middle) and 10 (bottom). Green = RANSAC inliers, red = outliers. FLANN: fast library for approximate nearest neighbors, RANSAC: random sample consensus, SP: SuperPoint.

Fig. 6 : Top-down stereo VO trajectories, SE(3)-aligned to GT, of the lowest (solid) and highest (dashed) ATE combination that tracks the whole path, on five datasets covering the three tiers. The S3LI sites show their scored first 10 %. ATE: absolute trajectory error, FLANN: fast library for approximate nearest neighbors, GT: ground truth, SP: SuperPoint, VO: visual odometry.

Panel of Figure 12 — Fig. 7 : Loop-closure examples. (a) Matches of SIFT+FLANN (left) and SuperPoint+LightGlue (right), with RANSAC inliers (green), outliers (red) and rotation and translation-direction errors $e_{R}$ , $e_{t}$ . The pairs are a same-direction revisit on LuPNT (base) and, on LuSNAR (5), a pair seen from directions $123^{\circ}$ apart that shares $40\,\%$ of the view. (b) S3LI (Etna) DINOv2 retrieval: one query whose top-1 is a true revisit (left) and one whose top-1 is a false proposal (right). Top-down maps at far right. FLANN: fast library for approximate nearest neighbors, RANSAC: random sample consensus, SP: SuperPoint.

Fig. 8 : Incremental cost per frame [ms] on LuPNT (base). Hatched bars run on the GPU. CPU: central processing unit, FLANN: fast library for approximate nearest neighbors, GPU: graphics processing unit.

Planetary Rover Robotics

See all 16 papers

Mar 18, 2026

Visual SLAM with DEM Anchoring for Lunar Surface Navigation

Future lunar missions will require autonomous rovers capable of traversing tens of kilometers across challenging terrain while maintaining accurate localization and producing globally consistent maps. However, the absence of global positioning systems, extreme illumination, and low-texture regolith make long-range navigation on the Moon particularly difficult, as visual-inertial odometry pipelines accumulate drift over extended traverses. To address this challenge, we present a stereo visual simultaneous localization and mapping (SLAM) system that integrates learned feature detection and matching with global constraints from digital elevation models (DEMs). Our front-end employs learning-based feature extraction and matching to achieve robustness to illumination extremes and repetitive terrain, while the back-end incorporates DEM-derived height and surface-normal factors into a pose graph, providing absolute surface constraints that mitigate long-term drift. We validate our approach using both simulated lunar traverse data generated in Unreal Engine and real Moon/Mars analog data collected from Mt. Etna. Results demonstrate that DEM anchoring consistently reduces absolute trajectory error compared to baseline SLAM methods, lowering drift in long-range navigation even in repetitive or visually aliased terrain.

Adam Dai, Guillem Casadesus Vila, Grace Gao

Stanford University William F. Durand Building 496 Lomita Mall Stanford, CA 94305

Simultaneous Localization and Mapping Visual SLAM Robot Navigation Map-Based Localization

Ask PDF

Aug 19, 2026

Evaluation of Image Matching Methods for Visual Odometry on UAVs

Unmanned aerial vehicles (UAVs) are becoming a powerful tool for many environmental monitoring and transport applications. Yet, their reliance on Global Navigation Satellite System (GNSS) technology for navigation makes them susceptible to catastrophic failures in scenarios where the positioning signal is unavailable or disrupted. This work explores Visual Odometry (VO) as a crucial navigation component. Recently, numerous deep-learning-based methods for image matching have been proposed that are yet to be implemented in a fully-fledged VO system. In this paper, we evaluate recent state-of-the-art image matching methods for the task of VO for UAV position tracking, with a downwards-facing camera, on our synthetic dataset, and find that while the best results are generated by the recent RoMa matcher, SIFT features can outperform some recent state-of-the-art.

Gašper Spagnolo, Luka Čehovin Zajc, Matej Dobrevski

Faculty of Computer and Information Science, University of Ljubljana, Večna pot 113, Ljubljana, Slovenia.

Visual Odometry Image Matching UAV Navigation

Ask PDF

Jul 2, 2026

DL-VINS-Factory: A Modular Framework for Learned Visual Front-Ends in Visual-Inertial SLAM

Deep-learning features excel in visual matching, yet their practical value in tightly coupled visual-inertial SLAM (VI-SLAM) remains insufficiently characterized. We present DL-VINS-Factory, a unified framework that integrates learned feature extractors (ALIKED, RaCo, SuperPoint, XFeat) with either Lucas--Kanade (LK) optical-flow tracking or LightGlue (LG) descriptor matching. All front-ends share a sliding-window Ceres back-end, with optional AnyLoc DINOv2-VLAD loop closure, and 4-DoF pose-graph optimization. We benchmark the system across the four datasets covering indoor, unstructured outdoor, aggressive-motion, and visually degraded conditions. Results show that learned front-ends are viable for real-time embedded VI-SLAM, but are not universally superior to classical tracking. Relative to the corresponding GFTT+LK baseline, ALIKED+LG reduces EuRoC ATE by

5\%

in monocular odometry and by

7\%

in stereo with loop-closure. On NTU-VIRAL, where aggressive aerial motion increases inter-frame viewpoint change, ALIKED+LG stereo reduces loop-closed ATE by

12\%

. In Botanic Garden dataset, optical-flow tracking remains preferable, but learned keypoints still improve over the baseline GFTT, in which SuperPoint+LK reduces grayscale camera ATE by

29\%

, while RaCo+LK reduces RGB camera ATE by

38\%

. On SubT-MRS, learned front-ends display varying degree of improvement based on individual cases. With TensorRT acceleration on a Jetson AGX Orin, all valid configurations run in real time between

29

47

FPS in monocular mode and

18

33

FPS in stereo mode for the EuRoC and NTU-VIRAL datasets. AnyLoc further confirms roughly

2

7\times

more valid loops than BRIEF+DBoW2. The implementation is open-sourced at https://github.com/limshoonkit/DL-VINS-Factory-ROS2/.

Shoon Kit Lim, Melissa Jia Ying Chong, Ting Yang Ling

University of Southampton Malaysia, Iskandar Puteri, Johor, Malaysia · University of Southampton, Southampton SO17 1BJ, U.K.

Visual-Inertial SLAM Visual-Inertial Odometry

Ask PDF

Abstract

Planetary Rover Robotics

Explore similar work

Visual SLAM with DEM Anchoring for Lunar Surface Navigation

Evaluation of Image Matching Methods for Visual Odometry on UAVs

DL-VINS-Factory: A Modular Framework for Learned Visual Front-Ends in Visual-Inertial SLAM