PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluation of Front-End Feature Matching for Lunar Visual SLAM

Guillem Casadesus Vila, Grace Gao

arXiv:2610.30774v1cs.CV

TL;DR

Feature matching between rover frames has not been compared, despite its role in camera-based lunar localization and visual odometry. The paper benchmarks detector–matcher combinations across nine datasets; SuperPoint performs better on lunar testbeds under rotation-only scoring, while results vary by scoring and task.

  • Problem

    Comparative evidence on feature matching between rover frames is lacking, although lunar rovers rely on cameras and visual odometry estimates motion by matching frames.

    Source: §1 · p. 1, §6 · p. 13

  • Method

    The paper benchmarks eleven detector–matcher combinations and a KLT tracker across nine datasets, measuring pairwise pose accuracy and resulting visual-odometry trajectory error.

    Source: §6 · p. 13

  • Results

    Rotation-only scoring favors SuperPoint on lunar testbeds, whereas joint scoring favors SIFT+FLANN on synthetic data; stereo VO shows no consistent accuracy advantage, but SuperPoint tracks all eight sequential datasets.

    Source: §6 · p. 13

  • Takeaways & Limitations

    For GPU-equipped rovers, the authors recommend SuperPoint with LightGlue; without a GPU, SIFT+FLANN has the highest classical Pose AUC at default settings.

    Source: §6 · p. 13

    The benchmark still needs extension to real lunar surface imagery from Chang’e-3 and Apollo.

    Source: §6 · p. 13

Abstract

from arXiv · show

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.

1. INTRODUCTION

Lunar rovers will depend on cameras for localization and other surface tasks, but matching their low-texture, harshly lit images across changing viewpoints remains challenging. This study benchmarks front ends for VO and loop closure across diverse datasets and evaluation conditions.

1. INTRODUCTION

  • Cameras will be the main localization sensor until LunaNet navigation satellite services are deployed and will remain essential for relative navigation, mapping, and hazard detection.The motivation includes planned rover drives far longer and faster than those of previous planetary rovers.Source: §1 · p. 1
  • Most VO and SLAM systems estimate relative pose by detecting and matching keypoints, so correspondence quality affects the trajectory and any resulting map.SLAM also bounds accumulated error when it recognizes revisited places.Source: §1 · p. 1
  • Lunar matching is difficult because regolith has little texture, optics add distortions and noise, and larger viewpoint changes arise when rovers drive faster or skip frames.Loop closure must match views of the same place from different positions and headings, a harder task where distinctive landmarks are scarce.Source: §1 · p. 1
  • Existing benchmarks often report one score per dataset, obscuring viewpoint-related degradation; synthetic-only results cannot establish transfer to real imagery, and retrieval and matching are often evaluated separately.SLAM needs both: retrieval proposes a revisit, and matching verifies it.Source: §1 · p. 1
  • The study benchmarks eleven detector–matcher combinations and a KLT tracker across nine datasets, evaluating pose error, computational cost, lighting effects, transfer from synthetic to real imagery, and VO and loop-closure relevance.It also characterizes operating range along the baseline and its correlation with ATE/L, and tests whether pairwise matching suffices for loop closure.Source: §1 · pp. 1–2

2. RELATED WORK

The related work reviews visual odometry and SLAM, feature detection and matching, space applications, matching benchmarks, and place recognition for loop closure. It situates the study’s camera-only front ends and evaluations within those areas.

Visual Odometry and SLAM

  • VO benchmarks score whole trajectories, while this study describes pair difficulty by overlap and triangulation angle; direct and end-to-end systems lack an exchangeable detector–matcher stage.KLT is identified as the closest direct-method baseline that yields correspondences; the evaluation focuses on the sparse front end.Source: Visual Odometry and SLAM · p. 2

Feature Detection and Matching

  • Classical methods hand-craft detection and description, whereas learned methods replace one or both steps; the study evaluates four classical detectors and SuperPoint with released weights for both SuperGlue and LightGlue.SuperGlue and LightGlue are used without retraining despite training on MegaDepth, so results include this domain gap; detector-free matchers and 3D foundation models are left out.Source: Feature Detection and Matching · p. 2

Feature Matching for Space Applications

  • Earlier space applications include stereo VO on Mars Exploration Rovers and orbital-map localization that complements VO; a simulated lunar trade study found crater localization degraded at low sun and with computed stereo depth.Dense intensity registration against orbital imagery stayed within 3 m in every case.Source: Feature Matching for Space Applications · p. 2

Benchmarking Feature Matching

  • The study adopts Pose AUC for matching quality, extends RUBIK’s overlap, scale-ratio, and triangulation-angle descriptions from one driving dataset to lunar and analogue data, and adds downstream VO and loop closure.Pairs are sampled on a grid of metric translations to reflect rover frame spacing.Source: Benchmarking Feature Matching · p. 2

Visual Place Recognition and Loop-Closure Evaluation

  • Prior lunar and planetary work does not evaluate the same camera-only front ends for both visual odometry and loop-closure verification across synthetic, lunar-testbed, and terrestrial-analogue data, or relate pairwise results to VO accuracy.The closest planetary-analogue loop-closure work, MPRF, combines DINOv2 retrieval with LiDAR-based pose estimation on S3LI sequences, including two used here.Source: Visual Place Recognition and… · p. 2

3. PROBLEM FORMULATION AND DATASETS

The section frames pairwise pose recovery as a function of viewpoint change and describes the datasets, ground-truth sources, and sampling used to evaluate it. It distinguishes overlap-based difficulty from translation and rotation, and cautions that viewpoint metrics are not directly comparable across dataset tiers.

Problem Formulation

  • The study asks which feature-matcher combinations recover relative pose most accurately on lunar and analogue imagery, how performance varies with baseline, and whether synthetic results transfer to real data and downstream tasks.The pairwise step detects and matches keypoints between frames before estimating relative pose; VO chains these estimates, while SLAM uses the same step for revisit verification.Source: Problem Formulation · p. 3

Describing Viewpoint Change

  • Viewpoint change is characterized primarily by overlap ratio, with median triangulation angle capturing parallax and scale ratio comparing co-visible-point distances; translation and rotation are not used as difficulty axes.Co-visibility uses valid nonsky depth and rejects occlusions by checking projected depth agreement within 5%.Source: Describing Viewpoint Change · p. 3
  • Reference-depth differences across dataset tiers make absolute overlap and triangulation-angle values incomparable, so accuracy is compared within tiers; S3LI depth uses RAFT-Stereo with a distance cutoff.On S3LI, the cutoff is Zmax = 0.1 fxB; on LuSNAR (5), cutting rendered depth closely reproduces stereo values.Source: Describing Viewpoint Change · p. 3

Datasets and Ground Truth

  • The nine datasets span synthetic LuPNT and LuSNAR, lunar testbeds PolarTraverse and LunaPolaris, and terrestrial-analogue S3LI.LuPNT provides a 378 m rover path under base, camera-effects, and high-sun conditions; the study uses LuSNAR scenes 9 and 5, downsampled to 512 by 512 pixels.Source: Datasets and Ground Truth · p. 3

PolarTraverse (Lunar Testbed)—

  • PolarTraverse uses 38 station pairs from a regolith-simulant testbed, while LunaPolaris provides rover sequences with lidar depth and motion-capture poses, including three lighting setups for comparison.PolarTraverse poses are refined with COLMAP and registered to a laser scan; LunaPolaris sequence 07 uses a low spotlight, and sequence 10 adds a rover headlight.Source: PolarTraverse (Lunar Testbed)— · pp. 3–4

S3LI (Terrestrial Analogue)—

  • S3LI contributes Etna and Vulcano sequences with GNSS positions but no dense ground-truth depth; its orientation is gyro-integrated, and epipolar precision, overlap, and triangulation angle are approximate.Its median per-pair rotation error is at most 0.31°, and results are checked under rotation perturbations up to three times that error.Source: S3LI (Terrestrial Analogue)— · p. 4

Pair Sampling

  • Pair sampling uses up to 1,000 independent pairs per dataset across frame steps with median translations of about 0.1–8 m; low-overlap pairs are absent from some datasets and common in others.Along traverses, translation, rotation, and parallax grow jointly, coupling these axes strongly in synthetic data.Source: Pair Sampling · p. 4, Figure 2 · p. 4

4. METHODS

The methods isolate front-end performance in open-loop trajectories, evaluating detector–matcher combinations and a feature tracker through pairwise geometry, visual odometry, revisit verification, and image retrieval.

4. METHODS

  • The evaluation uses the front end alone, with no pose-graph or bundle-adjustment back end, so trajectories are open-loop.Source: §4 · p. 4

Detectors and Matchers

  • The benchmark tests 11 detector–matcher combinations, plus four detector–matcher ablations that separate SuperPoint or RootSIFT from the matching method.Detectors are SIFT, ORB, AKAZE, BRISK, and SuperPoint; the 11 combinations pair the first four with FLANN and BruteForce, and add SIFT+LightGlue, SP+SuperGlue, and SP+LightGlue. FLANN uses a 0.75 ratio test, BruteForce uses mutual cross-check, and randomized FLANN results use the median of five seeds; a Shi–Tomasi/Lucas–Kanade tracker provides a baseline.Source: Detectors and Matchers · pp. 4–5

Visual Odometry Front-End

  • Frame-to-frame VO estimates pose with depth-backed PnP and chains accepted poses; stereo VO supplies disparity depth, while rendered ground-truth depth isolates matching and pose estimation on two datasets.If too few inliers have metric depth, a frame is matched against the last accepted frame. Stereo VO runs on sequential datasets; PolarTraverse has no image sequence and is evaluated pairwise only, and real versus synthetic VO results are not directly comparable.Source: Visual Odometry Front-End · p. 5

Evaluation Metrics and Protocol

  • Pairwise evaluation uses epipolar precision, reprojection precision, and Pose AUC@5°; pose scoring is rotation-only where reference translation direction is insufficiently accurate.Pose AUC uses RANSAC and a five-point essential-matrix solver; joint rotation-and-translation-direction error is used on synthetic data and PolarTraverse. Table 3 scores rotation alone across all nine datasets, and unestimable poses count as misses.Source: Evaluation Metrics and Protocol · p. 5
  • Trajectory evaluation aligns VO poses to ground truth without scale and reports ATE and ATE/L only for runs with at least 90% path coverage.Gaps longer than 10 frames count as lost, and stopped runs are marked rather than assigned ATE values. ATE/L is a whole-trajectory error and is not directly comparable to start-anchored rover end-point error.Source: Evaluation Metrics and Protocol · p. 5
  • Revisit verification scores pose recovery on true revisit pairs, while DINOv2 retrieval is evaluated at rover-scale radii using recall@1, AUPRC, and R@100P.Retrieval queries frames more than 20 m apart along the path, either causally or offline; the top match is correct when cameras are within the selected radius. Revisit pose recovery reports Pose AUC, candidate-match count, and RANSAC inlier fraction.Source: Evaluation Metrics and Protocol · p. 6

Comparing the Operating Range with VO ATE/L

Implementation

  • Experiments use published defaults without per-dataset tuning, with learned methods on GPU and classical methods on CPU.ORB is capped at 500 keypoints by default; relaxed ORB and AKAZE settings are also tested. Timing on consecutive LuPNT frames counts extraction and matching, but not stereo matching or pose estimation.Source: Implementation · p. 6

5. RESULTS AND DISCUSSION

Front-end performance varies with metric, dataset, baseline, and lighting, so synthetic rankings alone do not establish which matcher will work best on rover imagery. SuperPoint pairings are not consistently more accurate in stereo VO, but they maintain tracking across all eight datasets; DINOv2 retrieval proposals require geometric verification.

Accuracy Across Datasets

  • SuperPoint’s testbed lead stems mainly from detector and descriptor quality, not matcher choice or keypoint count; EP rewards precise localization and should not select a front end alone.Tuning ORB and AKAZE improves their low-texture results, and classical matching needs a ratio test or cross-check because RANSAC alone does not recover lost accuracy.Source: Accuracy Across Datasets · pp. 6, 8
  • Under joint pose AUC, SIFT+FLANN beats both SuperPoint pairings on four of five synthetic datasets, while SuperPoint leads every real dataset except S3LI (Vulcano); rotation-only rankings differ.Neither metric alone predicts localization performance, which requires both rotation and translation direction.Source: Accuracy Across Datasets · p. 6, Table 2 · p. 7, Table 3 · p. 7

Accuracy Versus Baseline

  • At the widest tested baselines, the best learned front end is about 2.5 times as accurate on S3LI (Etna) and nearly 8 times on LunaPolaris; SuperPoint operating range is roughly 2–3 times classical alternatives on three datasets.Short baselines poorly constrain translation direction, while long baselines lose overlap; chained KLT is strongest on synthetic datasets but cannot skip frames or verify revisits.Source: Figure 4 · p. 8, Table 4 · p. 8, Accuracy Versus Baseline · pp. 8–9

Camera Effects, Lighting and Cross-Dataset Transfer

  • On pose-matched LunaPolaris views, spotlight lighting significantly improves classical-detector combinations while SuperPoint shows no resolved change, consistent with added contrast and keypoints revealing terrain texture.Modeled camera effects lower EP but generally do not lower classical pose accuracy; the Etna–LuPNT comparison changes scene and geometry as well as appearance.Source: Camera Effects, Lighting and… · p. 9

Visual Odometry

  • Across eight datasets, only SuperPoint pairings keep tracking on every stereo-VO path, but they show no resolved trajectory-accuracy advantage over SIFT pairings.On LuPNT, SuperPoint stereo ATE is 1.7 to 3.3 times its ground-truth-depth ATE, consistent with stereo depth error limiting accuracy; ground-truth-depth VO otherwise barely separates front ends.Source: Visual Odometry · pp. 10–11, Table 6 · p. 11

Loop Closure

  • On same-direction LuPNT revisits, both SuperPoint pairings reach about five times SIFT+FLANN’s Pose AUC, while DINOv2 retrieval has R@100P below 0.07 in every row and needs verification.Position alone does not define a usable revisit: viewing direction and overlap matter, and widely different views often defeat pose recovery.Source: Table 7 · p. 12, Loop Closure · p. 12

Computational Cost

  • At 16 ms per frame, SP+LightGlue is the fastest learned combination and costs no more than most timed classical combinations on the tested hardware.It is about three times as fast as SuperGlue; rover-class processor performance was not tested.Source: Figure 8 · p. 13, Computational Cost · p. 13

6. CONCLUSION

The study benchmarks eleven detector–matcher combinations and a KLT tracker across nine synthetic, lunar-testbed, and terrestrial-analogue datasets, evaluating pairwise pose accuracy and VO trajectory error. Results depend on scoring metric and lighting, while SuperPoint pairings offer the most consistent tracking across sequential datasets.

6. CONCLUSION

  • Rotation-only scoring favors SuperPoint over SIFT+FLANN on both lunar testbeds and one analogue site, while joint scoring favors SIFT+FLANN on four synthetic datasets.Synthetic results alone therefore do not establish relative performance on real rover imagery; the testbed lead is attributed mainly to the SuperPoint detector and descriptor, not the learned matcher.Source: §6 · p. 13
  • Only the SuperPoint pairings maintain tracking on all eight sequential datasets, although VO accuracy advantages over other front ends are not consistent.About half of ground-truth-depth ATE differences fall within run-to-run variation, and stereo depth resolves no consistent SuperPoint advantage. Classical pairings improve under low spotlight versus diffuse lighting, while SuperPoint changes little; modeled camera effects lower EP but not classical Pose AUC. KLT is most accurate on synthetic datasets under the joint metric and slightly exceeds SuperPoint on LunaPolaris, but not Etna, and cannot skip frames or verify revisits.Source: §6 · p. 13
  • About five times the Pose AUC is reached by SuperPoint on same-direction retraces compared with SIFT+FLANN, but widely different revisit directions are almost never recovered.DINOv2 retrieval proposals still require geometric verification; SuperPoint may matter most on uniformly lit regolith and possibly in deep polar shadow.Source: §6 · p. 13
  • For GPU-equipped rovers, the authors recommend SuperPoint with LightGlue as the faster SuperPoint pairing, about as accurate as SuperGlue.Without a GPU, SIFT+FLANN has the highest classical Pose AUC at default settings, though lower-threshold AKAZE surpasses it on a LunaPolaris subsample; where SIFT+FLANN tracks, its VO is not consistently less accurate than SuperPoint's.Source: §6 · p. 13
  • Future work includes dataset-specific classical-detector tuning, simulation sweeps of sun elevation, geometry-based pair selection, and adapting learned matchers to analogue imagery.The authors also plan complete SLAM evaluation on overlap-defined revisits, a camera-independent S3LI reference, and benchmark extensions to Chang’e-3 and Apollo lunar imagery.Source: §6 · p. 13

BIOGRAPHY[

The section profiles two Stanford aerospace researchers whose work centers on space exploration and navigation.

BIOGRAPHY[

  • Guillem Casadesus Vila is a Stanford Ph.D. candidate whose research interests include robotic space exploration and space navigation and communication.He received a 2022 B.Sc. in Aerospace and Telecommunications Engineering from the Universitat Politècnica de Catalunya under the CFIS program.Source: BIOGRAPHY[ · p. 16
  • Grace Gao is a Stanford associate professor whose research focuses on robust and secure positioning, navigation, and timing, including applications to space robotics.Before Stanford, she was an assistant professor at the University of Illinois at Urbana-Champaign; she earned her Ph.D. at Stanford.Source: BIOGRAPHY[ · p. 16

Related papers

Loading 2610.30774v1…