{"id":"56b53652-a6a6-4c19-82d1-9ac6c2dd099d","arxiv_id":"2507.17130","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MARSCalib is a spherical-target LiDAR-camera extrinsic calibration method that adds corruption-robust ellipse detection and weighted pointcloud sphere fitting, and it reports strong results across three LiDAR types.","lead":"MARSCalib calibrates a camera and a LiDAR using a small sphere as the calibration target, and it is designed to keep working when the sphere or the camera lens gets dirty, blurred, scratched, or partially damaged. The authors test it with three different LiDAR types and several camera angles in a moon-like test field and a construction site, reporting lower calibration error than checkerboard, AprilTag, and other spherical-target baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table II contains exact duplicate translation and rotation errors across independent LiDAR-camera pairs, so the central quantitative claim of consistent superiority across sensor types is not currently supported.","rationale":"The reader's weakest_assumption focuses on Eq. (3), where the stated formula divides by the number of cells M rather than the total point count Σn_i. This is indeed an internal inconsistency between the text and the formula, and it should be corrected. However, it is most plausibly a typo in the write-up: if the implementation actually used Eq. (3) as written, the representative points would be scaled by the average cell occupancy and the sphere center estimate would be badly biased, which is inconsistent with the reported centimeter-level accuracy. The code is public, so this is easily checked. The more load-bearing concern is the exact duplication of numerical results in Table II across independent sensor configurations. This directly affects the paper's central quantitative claim and cannot be explained as a typo. Therefore I partially agree with the reader: Eq. (3) is a valid but secondary issue, while the duplicated table entries are the critical weakness. The appropriate verdict remains CONDITIONAL, because the concern is resolvable by checking the repository and regenerating the table; it does not, by itself, invalidate the method, but it does prevent acceptance of the current quantitative claims as reported.","tokens_in":11991,"tokens_out":4030,"duration_ms":42407,"concrete_test":"In the public repository (github.com/sparolab/MARSCalib), locate the recorded calibration results for the duplicated pairs: Ouster Cam2 vs Livox Cam2 and Ouster Cam3 vs Livox Cam3. Recompute translation, rotation, and reprojection errors from the raw extrinsics and ground-truth transforms. If the recomputed values for the two LiDARs are not identical within 0.0005 m and 0.0005°, the table entries are erroneous and all comparisons in Table II must be regenerated. Additionally, run the calibration pipeline at least 10 times per sensor configuration and report mean ± std to establish whether the claimed differences are significant.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The primary evidence for the paper's strongest claim (Section V: lower translation, rotation, and reprojection errors than [17]* and [42] across Ouster, SOSLAB, and Livox LiDARs) rests on Table II. That table contains exact numerical duplicates across physically independent configurations: for 'Ours', Ouster Camera 2 and Livox Camera 2 both report trans=0.018 m, rot=0.157°, and Ouster Camera 3 and Livox Camera 3 both report trans=0.028 m, rot=0.587°; for [17]*, Ouster Camera 2 and Livox Camera 2 both report trans=0.056 m, rot=0.889°. Different LiDARs have different scan patterns, noise characteristics, and extrinsics, so exact equality to three decimal places is not expected. This strongly suggests a data-handling, copying, or reporting error in the table. Because Table II is the quantitative foundation for the claimed superiority, the central claim is unsupported until these entries are verified and corrected. The paper also reports no variance or confidence intervals, so even where entries differ, statistical significance is unknown. This is not an allegation of misconduct; it is a concrete reason to require data provenance and re-computation before accepting the headline results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MARSCalib, a LiDAR-camera extrinsic calibration method that uses a small spherical target (10 cm radius) and is designed for multi-robot outdoor and planetary environments where both the target and the sensors may be contaminated or damaged. The camera pipeline uses SAM-based segmentation, Canny edge detection, an iterative ellipse-fitting pipeline with evaluation and rectification, and a perspective-error compensation step. The LiDAR pipeline uses Hough transform-based region detection, clustering along the beam direction with a frequency-weighted representative point selection, and sphere fitting via four-point combinations weighted by the known sphere radius. The extrinsic transform is obtained by PnP with robust optimization. Experiments are reported for three LiDAR types (Ouster OS1-32, SOSLAB ML-X, Livox MID-360) and three camera poses in a construction field and a simulated planetary environment, comparing against two target-based baselines and two targetless baselines, and evaluating robustness under target contamination and damage. The central claim is that the proposed method is more accurate and more robust than existing approaches across sensor types and under corruption.","tokens_in":12198,"tokens_out":4509,"duration_ms":48676,"significance":"If the claims hold, the paper makes a credible contribution to practical extrinsic calibration in field robotics: it addresses an underexplored scenario (simultaneous target and sensor corruption), introduces a small spherical target that can be carried by a robot, and offers a fully automatic detection pipeline. The comparison across three mechanically different LiDARs and the corruption experiments are valuable. The release of code is a positive step for reproducibility. However, the significance is dampened by the technical inconsistency in Eq. (3), the suspicious duplicate entries in Table II, the absence of error bars and statistical significance analysis, and the undisclosed thresholds in several algorithm steps. These issues currently prevent the reader from fully trusting the quantitative claims of superiority and robustness.","major_comments":[{"comment":"Equation (3) does not implement the frequency-weighted mean described in the text. The text states that a weighted sum based on frequency is used to select a representative point p_r, but the equation p_r = (Σ_{i=1}^M n_i c_i) / M divides by the number of cells M, not by the total number of points Σ n_i. As written, the formula weights cells equally rather than by point frequency, which contradicts the stated method and affects all subsequent sphere fitting and center estimation. Please correct the denominator or clarify the intended definition, and explain how this affects the implementation and the results.","section":"§3.2, Eq. (3)"},{"comment":"Table II contains exact duplicate numerical values across different LiDAR-camera configurations that are expected to be independent. For example, \"Ours\" reports trans=0.018 m and rot=0.157° for both Ouster Camera 2 and Livox Camera 2, and trans=0.028 m and rot=0.587° for both Ouster Camera 3 and Livox Camera 3; the [17]* column shows trans=0.056 m and rot=0.889° for both Ouster Camera 2 and Livox Camera 2. Since the three LiDARs have different scan patterns, noise characteristics, and extrinsics, exact equality to three decimal places is not physically expected. This strongly suggests a data handling, copying, or reporting error. Because Table II is the primary quantitative evidence for the claimed superiority over baselines, please verify all entries, correct any errors, and report the underlying per-scene statistics (or provide a clear explanation if these values are not independent measurements).","section":"§4.3, Table II"},{"comment":"Several load-bearing parameters and thresholds are not reported, which prevents reproducibility and a full assessment of robustness: the length threshold used to remove noisy LiDAR clusters, the validity threshold in the ellipse rectification step (what portion of P_s must lie outside the re-fitted ellipse for the ellipse to be deemed invalid), the Hough transform parameters for sphere region segmentation, the voxel size used for non-repetitive LiDAR representative point extraction, and the outlier rejection threshold in the PnP optimization. The availability of code is helpful, but the paper should state these values or describe how they were chosen, especially because the method's claimed robustness depends on them.","section":"§3, overall method"},{"comment":"The experimental claims of accuracy and robustness are reported only as single-point estimates (likely means) without error bars, standard deviations, or any statistical significance test. Table II is based on 10 scenes per configuration and Table III on 8 scenes, so per-scene variance can be computed and should be reported. Without this information, the reader cannot determine whether the differences between the proposed method and the baselines are meaningful, especially in cases where the reported values are close (e.g., Table III, C, Cam1-OS). Please provide mean ± std. dev. for all reported metrics or an equivalent statistical summary.","section":"§4.3 and §4.4, Tables II and III"},{"comment":"The ellipse rectification step appears to be stochastic: the text states that a \"single point randomly drawn from P_s ∩ P_c^e\" is used in the iterative re-fitting. This introduces run-to-run variability in the detected ellipse center, which is not analyzed or discussed. Please specify the number of random trials, the random seed policy, or provide a deterministic variant, and quantify the sensitivity of the final calibration result to this randomness.","section":"§3.1, Ellipse Rectification"}],"minor_comments":[{"comment":"The text claims the spherical target \"consistently achieved high detection rates in all conditions,\" but Table I shows that under extreme contamination at 45° and 60° the sphere is marked with a triangle (detected in only one rotation direction). Please qualify this statement or define the metric for \"high detection rate\" more precisely.","section":"§4.2, Table I"},{"comment":"The three contamination levels (easy, medium, extreme) are not defined quantitatively. Please specify the exact procedures or amounts of contamination used for each level so that the detection experiments are reproducible.","section":"§4.2"},{"comment":"The baseline [17]* is described as a modified version of Tóth et al. because it is \"not well-suited for detecting small spheres\" and is provided with a SAM-derived mask and a cropped sphere region. These modifications should be described in more detail so that the comparison is fair and reproducible, and so the reader can judge how much of the performance difference is attributable to the proposed method versus the baseline's adaptation.","section":"§4.3"},{"comment":"The claim of using \"the smallest spherical target to date\" is not supported by a quantitative comparison with the sphere sizes used in the cited works [17], [18], [19], [20]. Please provide a brief comparison or cite the relevant radii.","section":"§2"},{"comment":"The notation \"HG\" is used both for a segment in Figure 4 and for \"half of the major axis length\" in the text. This is confusing; please use a different symbol for the half-major-axis length or clarify the notation.","section":"§3.1, Eq. (1)"},{"comment":"The conclusion states that \"our experiments demonstrate that ... the spherical target remains reliably detectable,\" but as noted above, Table I shows some detection failures at extreme contamination and high angles. Please align the conclusion with the actual data.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The two most significant concerns are the inconsistency in Eq. (3) and the exact duplicate entries in Table II, which appear for different LiDAR models. These issues directly affect the paper's central claims of accuracy and cross-sensor consistency. If the duplicate entries turn out to be errors, the corrected table may change the conclusions. The lack of error bars and the undisclosed thresholds are also serious but more standard to address in a revision. I recommend major revision and, importantly, request the authors to provide the per-scene raw data or a clear provenance for the reported numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zed, quick take on MARSCalib. The core idea is a good one: a 10 cm sphere, detected with SAM and a custom ellipse rectifier in the camera, and a frequency-weighted point selection in the LiDAR, giving automatic extrinsic calibration that survives dirt, blur, and damage. The field work is unusually broad—three LiDAR types, three camera angles, real contamination, and a planetary test site. If the numbers are right, this is a useful advance for field robotics.\n\nWhat's genuinely new is the handling of corrupted sphere images: SAM segmentation, iterative ellipse rectification that excludes points from damaged regions, and a perspective-center compensation formula. The LiDAR side's hierarchical weighted sum on accumulated clusters is a sensible answer to the noise you get on curved surfaces. They also benchmarked against an independent checkerboard ground truth and release code. That is more than half the subfield does.\n\nThe problems are in the reporting. Table II contains exact duplicate translation and rotation values for Ours across Ouster Camera 2 and Livox Camera 2 (0.018 m, 0.157°) and across Ouster Camera 3 and Livox Camera 3 (0.028 m, 0.587°), and the same duplicates appear for [17]*. Different LiDARs do not produce identical numbers to three decimals. This looks like a data-handling error and voids the claim of consistent superiority until the true numbers are provided. Also, Eq. (3) gives pr = (Σ n_i c_i)/M, but the text says frequency-weighted mean—it should be divided by Σ n_i, not M. Probably a typo, but it makes the method description inconsistent. No error bars are reported anywhere; with 10 scenes per configuration, standard deviations or medians should be presented. Several thresholds (cluster length, ellipse validity, outlier rejection) are hand-set but not listed, which hurts reproducibility despite the code link.\n\nWho this is for: researchers doing LiDAR-camera calibration in outdoor, dusty, or harsh environments. It deserves a serious referee, but only after the authors re-verify the Table II entries and fix the equation. I would not desk-reject it; I'd send it back for a careful revision with a request for raw data or a corrected table.","headline":"Promising corruption-robust sphere calibration pipeline, but duplicate numbers in Table II undermine the central comparative claim until corrected.","tokens_in":12746,"tokens_out":2494,"would_cite":false,"duration_ms":24198,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a 10-cm-radius styrofoam sphere is enough to automatically recalibrate LiDAR-camera extrinsics in outdoor and planetary multi-robot deployments, even when the target is contaminated or damaged and the camera is…","keywords":["LiDAR-camera extrinsic calibration","spherical calibration target","multi-robot calibration","segmentation","pointcloud denoising","robust target detection","planetary exploration","sensor contamination"],"falsifier":"Run a clean-laboratory reproduction of Table II with a millimeter-accurate checkerboard reference and also compute Eq. (3) on a synthetic cluster with known cell occupancies; the formula as written does not equal the frequency-weighted mean described in the text, so if the implementation follows the paper, the fitted sphere center—and hence the reported translation errors—will shift by an amount that a precise ground-truth check will reveal.","tokens_in":11754,"feed_emoji":"🎯","tokens_out":7868,"duration_ms":76897,"temperature":0.7,"pith_summary":"This paper tries to establish that a small, robot-portable sphere can serve as a reliable LiDAR-camera calibration target in unstructured outdoor and planetary settings, where conventional checkerboards and fiducial markers fail under mud, blur, scratches, and partial damage. The proposed method combines SAM-based image segmentation with an iterative ellipse-rectification routine and an analytical compensation for perspective projection error, so the 2D ellipse center faithfully tracks the sphere's visual center. On the pointcloud side, it claims that a hierarchical weighted-sum of accumulated LiDAR measurements, after removing long noise clusters, yields unbiased surface points for sphere fitting, making the approach work across spinning, solid-state, and non-repetitive LiDARs. The paper's experiments report lower translation, rotation, and reprojection errors than two target-based baselines on three sensor pairs, and higher detection rates than AprilTag and CopperTag under contamination. If true, the payoff is fully automatic, human-free recalibration for multi-robot fleets operating in the field.","feed_headline":"One 10-cm sphere calibrates LiDAR-cameras even when dirty","feed_subtitle":"Spherical target survives mud, blur, and scratches, so robot fleets can recalibrate themselves without a human in the loop.","key_machinery":"The argument is carried by two mechanisms plus a correction formula. First, the ellipse rectification loop: starting from Canny edges on SAM masks, the method fits an initial ellipse, classifies it by the angular distribution of sampled perimeter points, and if the distribution is concentrated, refits using a random point from outside the current ellipse together with points from each concentrated region, iterating until the ellipse is stable. This loop is what makes the method tolerant to target corruption, because intact regions of the sphere perimeter dominate the final fit. Second, the noise-aware LiDAR center extraction: clustered points from the same surface location are reduced to a single representative point by a hierarchical weighted sum; it is this representative-point set, not the raw noisy sphere points, that feeds the sphere-fit and center-combination stage. Third, the center-compensation formula in Eqs. (1)-(2) corrects the systematic error that arises because a sphere's projected ellipse center does not coincide with the projection of the sphere center unless the sphere is centered on the optical axis; the correction is applied before the 2D point is paired with the 3D center.","core_discovery":"On the paper's own terms, the central discovery is that a 10-cm-radius styrofoam sphere, the smallest spherical calibration target used to date, is sufficient for accurate LiDAR-camera extrinsic calibration, provided the two sensing modalities are processed with corruption-aware steps. In the image, the sphere's projection is an ellipse; the method first segments the image with SAM, then fits an ellipse while iteratively refitting against points outside the current ellipse to avoid the bias introduced by damaged or contaminated regions, and finally shifts the ellipse center using a closed-form compensation for the perspective-projection error between the ellipse center and the sphere's true projected center. In the LiDAR pointcloud, the method accumulates static data, removes clusters elongated along the laser ray that indicate noise, subdivides surviving clusters into cells, and selects one representative point per cluster via a frequency-weighted average; sphere centers are then fitted from four-point combinations and combined by frequency. The claim is that these two center estimates, when paired through a PnP solver with a robust kernel, give state-of-the-art calibration accuracy across three LiDAR types and three camera viewpoints, and remain reliable even when the sphere is dirt-covered, truncated by about 25%, or scratched and the camera is blurred or mud-spattered.","pith_inferences":["The center-compensation step, if as accurate as claimed, applies beyond calibration: any monocular system that needs the unbiased image-center of a spherical object—visual tracking, sphere-based robotic manipulation—could reuse the same correction formula.","The method's reliance on SAM for initial segmentation is not evaluated for compute cost or latency; a lighter-weight segmentation network might preserve the rectification benefits while running on onboard robot computers, but that is a testable extension the paper does not report.","The equation for the frequency-weighted representative point appears inconsistent with the surrounding text: the formula divides by the number of cells rather than the total point count, so the reported accuracy could depend on an undocumented weighting scheme if the printed formula is what the code implements.","The corruption experiments cover contiguous damage (dirt, scratches, a single 25% truncation); a sphere with many small widely scattered dents might violate the 'concentrated region' assumption in the ellipse rectification, and that case is not tested."],"forward_implications":["A robot team can trigger recalibration on demand: one robot holds up the small sphere while a drifted LiDAR-camera pair observes it, and the full pipeline runs without human intervention.","The same segmentation and rectification pipeline works for spinning, solid-state, and non-repetitive LiDARs, so a single calibration procedure can cover a heterogeneous fleet.","Because the sphere is only 10 cm in radius and made of styrofoam, it can be carried by a manipulator or mounted on a robot, making calibration practical in field conditions where large boards cannot be transported.","Under target and sensor contamination, the sphere remains detectable in cases where AprilTag and CopperTag fail, so calibration can be maintained in muddy, dusty, or blurry conditions.","The reported error levels (translation around 0.01–0.04 m, rotation below 1 degree in most configurations) suggest the method meets the accuracy needed for downstream LiDAR-camera fusion tasks such as SLAM and object detection."],"supporting_citations":[{"why":"Baseline spherical-target calibration method that the paper adapts into [17]* and reports lower errors against.","marker":"[17]"},{"why":"Baseline tag-and-reflective-tape target method that provides the second target-based comparison in Table II.","marker":"[42]"},{"why":"Segment Anything Model, whose masks narrow the search region and make small-sphere detection possible in the camera pipeline.","marker":"[21]"},{"why":"CopperTag, the occlusion-resilient fiducial marker used as a robustness comparison under contamination.","marker":"[30]"},{"why":"AprilTag, the widely used fiducial marker used as a robustness comparison under viewing-angle and contamination.","marker":"[41]"},{"why":"Levenberg-Marquardt optimizer used to solve the PnP problem for the extrinsic transformation.","marker":"[35]"},{"why":"Statistical Outlier Removal from the Point Cloud Library, applied to filter LiDAR data before sphere extraction.","marker":"[38]"},{"why":"MATLAB LiDAR-camera calibration toolbox used with a large checkerboard to establish ground truth.","marker":"[40]"}],"fun_headline_variants":["Smallest sphere yet calibrates LiDAR-camera in mud and scratches","Automatic sphere calibration for robots in dirt and dust","Calibrate multi-robot sensors with a tiny scuffed sphere","Dirty, scratched sphere still gives precise LiDAR-camera calibration","Spherical target corruption-proofs LiDAR-camera calibration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that LiDAR points on a sphere can be reduced to unbiased surface samples by removing long noise clusters and averaging each remaining cluster, and that the printed formula for this average—which divides by the number of cells, not the number of points—actually produces the intended frequency-weighted mean.","fun_headline_variants_meta":{"raw":{"variants":["Smallest sphere yet calibrates LiDAR-camera in mud and scratches","Automatic sphere calibration for robots in dirt and dust","Calibrate multi-robot sensors with a tiny scuffed sphere","Dirty, scratched sphere still gives precise LiDAR-camera calibration","Spherical target corruption-proofs LiDAR-camera calibration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000749,"raw_usage":{"total_tokens":3395,"prompt_tokens":1067,"completion_tokens":2328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":683,"completion_tokens_details":{"reasoning_tokens":2242}},"tokens_in":683,"tokens_out":2328,"duration_ms":16929,"temperature":1.0,"reasoning_tokens":2242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:55:58.368780+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a clean-laboratory reproduction of Table II with a millimeter-accurate checkerboard reference and also compute Eq. (3) on a synthetic cluster with known cell occupancies; the formula as written does not equal the frequency-weighted mean described in the text, so if the implementation follows the paper, the fitted sphere center—and hence the reported translation errors—will shift by an amount that a precise ground-truth check will reveal.","supporting_citations":[{"cited_title":"Automatic lidar-camera calibration of extrinsic parameters using a spherical target,","cited_arxiv_id":null,"evidence_quote":"Baseline spherical-target calibration method that the paper adapts into [17]* and reports lower errors against."},{"cited_title":"An effective camera-to-lidar spatiotemporal cali- bration based on a simple calibration target,","cited_arxiv_id":null,"evidence_quote":"Baseline tag-and-reflective-tape target method that provides the second target-based comparison in Table II."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Segment Anything Model, whose masks narrow the search region and make small-sphere detection possible in the camera pipeline."},{"cited_title":"Coppertag: A real-time occlusion-resilient fiducial marker,","cited_arxiv_id":null,"evidence_quote":"CopperTag, the occlusion-resilient fiducial marker used as a robustness comparison under contamination."},{"cited_title":"Apriltag: A robust and flexible visual fiducial system,","cited_arxiv_id":null,"evidence_quote":"AprilTag, the widely used fiducial marker used as a robustness comparison under viewing-angle and contamination."},{"cited_title":"The levenberg-marquardt algorithm,","cited_arxiv_id":null,"evidence_quote":"Levenberg-Marquardt optimizer used to solve the PnP problem for the extrinsic transformation."},{"cited_title":"3d is here: Point cloud library (pcl),","cited_arxiv_id":null,"evidence_quote":"Statistical Outlier Removal from the Point Cloud Library, applied to filter LiDAR data before sphere extraction."}],"review_version":1}