{"id":"b027603a-464b-41cd-8252-a8249f2ec57b","arxiv_id":"2507.17210","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A target-based, automatic LiDAR-camera extrinsic calibration pipeline detects circular hole centers from any LiDAR scan pattern and solves the rigid transform in under 0.7 seconds.","lead":"FAST-Calib is an automatic tool that aligns a LiDAR and a camera by detecting four circular holes on a printed board and computing the sensor transformation in under one second. It is aimed at robotics teams that need fast, repeatable calibration for both mechanical and solid-state LiDARs, and the code and dataset are public.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sub-6.5 mm accuracy claim is supported only by the least-squares residual that the extrinsic is fitted to minimize, so it does not independently validate the extrinsics.","rationale":"The reader's verdict of CONDITIONAL is appropriate, and my concern does not move it. I partially agree with the reader's weakest-assumption identification: biased LiDAR-side hole centers would corrupt the correspondences. However, I would sharpen the primary issue: the paper's accuracy evidence cannot detect such bias because it is measured with the same objective minimized during calibration. Even a method with systematically biased hole extraction could report sub-millimeter residuals, as Table I does. The concrete fix is a held-out or externally grounded accuracy evaluation. The runtime claim is well supported by Table II, and the consistency scatter plots support repeatability, but repeatability and self-residual do not establish extrinsic accuracy. The open-source release is a genuine strength and makes the proposed held-out experiment easy to run. The paper is a solid engineering contribution, but the headline accuracy claim needs independent validation before full acceptance.","tokens_in":6973,"tokens_out":2619,"duration_ms":31051,"concrete_test":"For each sensor configuration, hold out one of the four collected data pairs. Calibrate on the remaining three pairs, then evaluate the held-out pair: compute the Eq. (5) point-to-point residual for the held-out 3D-3D hole correspondences, and also project the LiDAR hole centers into the camera image using the fitted extrinsic to compare against image-derived hole positions. Report held-out residuals separately from the in-sample residuals in Table I. If held-out errors remain below 6.5 mm and are comparable to in-sample values, the circularity concern is resolved; if they are substantially larger, the stated accuracy claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core accuracy evidence is Table I, which reports residuals of the objective in Eq. (5) after Kabsch registration. These residuals are computed on the same point correspondences used to estimate T_CL, so low values only show that the fitted transform is self-consistent with the extracted hole centers; they do not show that the hole centers or the resulting extrinsics are correct. If the angular-gap edge detector or ellipse fitter in Eqs. (1)-(4) returns systematically biased LiDAR hole centers, or if the ArUco-derived camera hole centers are biased by a small PnP pose error, the Kabsch solution will absorb much of that bias and still produce small residuals. The comparison with Velo2Cam in Table I uses the same residual metric, so it is also circular with respect to absolute accuracy. The abstract's 'registration errors consistently below 6.5 mm' is therefore a self-consistency figure, not an independent accuracy figure. Without a held-out scene, an external ground-truth extrinsic, or a reprojection check on data not used in the fit, the central accuracy claim is not established. The open-source code is a positive point, but the validation gap remains the load-bearing concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FAST-Calib is a target-based LiDAR-camera extrinsic calibration pipeline. It uses a custom board with four circular holes and four ArUco markers. In the camera branch, ArUco pose estimation via PnP places the hole centers in the camera frame. In the LiDAR branch, pass-through ROI filtering, RANSAC plane segmentation, 8 mm voxel downsampling, and an angular-gap edge detector (Eqs. 1-2) extract hole-edge points, which are grouped and fit with ellipses (Eqs. 3-4) to compensate for spot-spread dilation; the ellipse centers give the hole centers in the LiDAR frame. The extrinsic transform is computed by Kabsch registration minimizing Eq. (5), extended to multiple captures. Experiments on Ouster OS1-128, Livox Avia, and Mid360 with a wide-angle camera report low residual values in Table I, consistency across data subsets in Fig. 3, and total runtime under 0.7 s in Table II. The code and dataset are open-sourced.","tokens_in":7252,"tokens_out":5496,"duration_ms":56312,"significance":"If the accuracy claim survives independent validation, FAST-Calib would be a practically valuable contribution: automatic, pattern-agnostic edge extraction that works for both mechanical and solid-state LiDARs; ellipse-based compensation for spot-spread dilation; joint multi-scene optimization; and very fast processing. The paper also releases code and data, which is a strength. However, the current evidence does not establish the central 'registration errors below 6.5 mm' claim, because the reported residuals are the objective being minimized; the significance therefore hinges on adding independent accuracy evaluation.","major_comments":[{"comment":"The residuals reported in Table I are computed from Eq. (5), the exact least-squares cost minimized by the Kabsch registration, using the same point correspondences that produced the transform. Low values (0.10-0.65 cm) therefore demonstrate self-consistency between the extracted LiDAR and camera hole centers, not absolute accuracy. Any systematic bias shared by the LiDAR-side edge/ellipse extraction and the camera-side PnP pose (for example, a dilation-compensation bias or an ArUco pose bias) will be partially absorbed by the fitted transform while still yielding small residuals. The abstract's 'below 6.5 mm' is thus a fitting residual, not an independent accuracy figure. I ask for a held-out evaluation: leave-one-capture-out reprojection of LiDAR points into images, comparison against an independent ground-truth extrinsic (e.g., hand-eye or mechanical reference), or residuals computed on data not used in the fit.","section":"IV.B / Table I / Eq. (5)"},{"comment":"The paper does not validate the assumption that the maximum-angular-gap rule with a 25 degree threshold and the subsequent ellipse fit recover true geometric hole centers for all three scan patterns. For sparse and irregular patterns (Mid360 in particular), edge arcs can be short and incomplete, making ellipse centers sensitive to missing data and to spot-spread dilation that may not be uniform. Since Table I cannot detect such bias, the paper should report hole-center extraction accuracy against the known physical board geometry (the hole radius and relative positions are known from Fig. 1) or compare LiDAR-extracted centers with centers obtained from a dense, high-resolution scan. Without this, the accuracy claim is not load-bearing.","section":"III.C / Eqs. (1)-(4)"},{"comment":"The comparison with Velo2Cam is not fully interpretable because the paper states that Velo2Cam was 'adapted' to Livox Avia and Mid360 but does not describe how the ring-based depth-discontinuity detection is defined for non-repetitive or sparse scan patterns. The reader cannot tell whether Velo2Cam's large residuals on solid-state LiDARs reflect the method's known limitation or a non-fair adaptation. Please specify the adaptation and, if possible, include failure modes/criteria.","section":"IV.B"},{"comment":"The reported 'total processing time' starts after data acquisition and accumulation; for sparse LiDARs the paper itself says accumulated scans may need FAST-LIVO2 odometry with pitch movements to densify the cloud. The title and abstract claim 'one second' calibration, but the time spent collecting and accumulating the point cloud is not counted. Please state explicitly that the runtime covers only the calibration computation after data collection, and if the claim is end-to-end, add the data-collection time.","section":"IV.C / Table II"}],"minor_comments":[{"comment":"Equation (5) is typeset with a broken fraction; please fix the 1/(4N) notation so the normalization is clear.","section":"III.D / Eq. (5)"},{"comment":"Table I lists residuals in centimeters while the abstract quotes millimeters; state the conversion explicitly (e.g., 0.15 cm = 1.5 mm) to avoid confusion.","section":"Table I / Abstract"},{"comment":"The text says the average pose of the four detected ArUco markers is adopted, but averaging poses is not well-defined for rotations; specify whether you average translations and rotations separately (e.g., quaternion averaging) or select one marker.","section":"III.B"},{"comment":"The statement that 16-line mechanical LiDARs may require FAST-LIVO2 odometry with pitch movements is not experimentally validated, since the evaluated sensors are OS1-128, Avia, and Mid360; please clarify that this is a recommendation rather than a validated procedure.","section":"III.A"},{"comment":"The key physical dimensions of the board and the hole/marker layout appear only in the technical drawing; listing them in the text or a table would make the method easier to reproduce.","section":"Fig. 1 / III.B"},{"comment":"The random x-offset added to the scatter points makes the visual spread in the x-direction partly artificial; consider using a small, fixed jitter or a box plot so the reader can judge the actual variance.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"I agree with the reader's assessment: the central accuracy claim needs non-circular validation. The method itself is coherent, the open-source release is a positive point, and the consistency and runtime evidence are useful. The paper is likely acceptable after adding a held-out or ground-truth accuracy evaluation and tightening the runtime and comparison claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: FAST-Calib is a genuinely practical target-based calibration tool, open-sourced, and it runs in under a second. The accuracy headline, however, is not backed by the experiments as written. The sub-6.5 mm numbers are residuals of the same least-squares objective the extrinsic is fitted to minimize, so they demonstrate self-consistency, not absolute accuracy.\n\nWhat's new: the combination of scan-pattern-agnostic edge detection via maximum angular gap, ellipse fitting to compensate for spot dilation, and multi-scene joint optimization. Each piece is standard—the board is from Velo2Cam, Kabsch is textbook—but the specific pipeline for solid-state LiDARs is new, and the paper is honest about reusing the target. The runtime breakdown is clear, and the consistency experiments across data combinations show low variance. Open-sourcing code and data is a real plus.\n\nWhere it's soft: the evaluation. Table I reports Eq. (5) residuals computed on the same 3D-3D correspondences used to estimate T_CL. Low residuals only show the fit converged. If the LiDAR hole centers are biased—by spot dilation, sparse beams, or PnP error on the ArUco side—Kabsch absorbs much of that bias and the residuals stay small. The Velo2Cam comparison uses the same metric, so it is also not an independent accuracy check. There is no held-out scene, no known ground-truth extrinsic, no reprojection on data not used in the fit. The dilation compensation via ellipse fitting is a claimed contribution but there is no ablation isolating it. Minor: for sparse mechanical LiDARs they fall back to FAST-LIVO2 odometry with slight motion, which is a reasonable practical choice but adds a dependency and possible drift.\n\nThe stress-test note is right: the central accuracy claim is not established. That said, the method itself is coherent and the engineering is credible. The reader's CONDITIONAL verdict with moderate confidence is fair.\n\nWho this is for: robotics practitioners who need quick, automatic target-based calibration for solid-state or mechanical LiDAR-camera rigs. They will read it for the pipeline and the code. A researcher in calibration evaluation will find the validation gap instructive.\n\nRecommendation: send it to peer review, but require major revision on the evaluation: an external accuracy assessment or held-out reprojection, and an ablation for the dilation compensation. The paper is worth a serious referee.","headline":"FAST-Calib is a fast, open-source calibration pipeline with a real practical niche, but its sub-6.5 mm accuracy claim is a fitting residual, not an independent validation.","tokens_in":7757,"tokens_out":2565,"would_cite":false,"duration_ms":26708,"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":"FAST-Calib claims LiDAR-camera extrinsics can be estimated in under one second with point-to-point residuals below 6.5 mm.","keywords":["LiDAR-camera calibration","extrinsic calibration","target-based calibration","solid-state LiDAR","mechanical LiDAR","angular gap edge extraction","ellipse fitting","Kabsch registration"],"falsifier":"Use a board with machined hole centers at known coordinates, with the LiDAR-camera transform independently measured (e.g., by a motion-capture reference), and vary the board's incidence angle. If the estimated extrinsics drift by more than the reported 6.5 mm residual when spot dilation is large or holes are viewed near tangentially, the assumption that fitted ellipse centers equal true hole centers is falsified.","tokens_in":6763,"feed_emoji":"🎯","tokens_out":7934,"duration_ms":75679,"temperature":0.7,"pith_summary":"This paper is trying to establish that target-based LiDAR-camera extrinsic calibration can be automatic, sub-second, and equally usable with mechanical and solid-state LiDARs. The proposed pipeline uses a board with four circular holes and four ArUco markers: hole edges are detected in the LiDAR point cloud by looking for large angular gaps in each point's local neighborhood, then an ellipse is fitted to each hole's edge points to undo the dilation caused by LiDAR spot spread. The fitted ellipse centers are registered against hole centers derived from the ArUco-detected board pose, and the transform is solved in closed form with Kabsch. The paper reports point-to-point registration residuals below 6.5 mm and total processing time under 0.7 s on Ouster, Livox Avia, and Mid360 configurations.","feed_headline":"Calibrate LiDAR-camera in one second with sub-6.5 mm error","feed_subtitle":"A board with four holes and ArUco markers works on mechanical and solid-state LiDARs, canceling spot-spread artifacts.","key_machinery":"The load-bearing object is a flat calibration board with four circular holes and four ArUco markers at known geometric positions. The mechanism that makes it scan-pattern agnostic is the angular-gap edge test: for each 2D point, compute the angles to all neighbors within a 3 cm radius (Eq. 1), sort them, and flag the point as an edge point if the largest gap between consecutive angles exceeds 25 degrees (Eq. 2). Edge points are clustered, and each cluster is fit with a general conic constrained to an ellipse (Eq. 3) using direct least squares; the analytic center (Eq. 4) is taken as the hole center, which compensates for edge dilation from LiDAR spot spread. Camera-side hole centers come from the ArUco markers' board pose, and the final transform is the closed-form Kabsch solution of the least-squares problem (Eq. 5), which naturally extends to $N$ captures by treating all $4N$ correspondences at once.","core_discovery":"The central claim is that a single target-based pipeline can deliver accurate extrinsic calibration in well under one second across scan-pattern families that previously required different tools. On the LiDAR side, the method avoids scan-line assumptions by classifying points as edge points from angular gaps in a 2D neighborhood, with a 25-degree threshold, and then replacing each circular hole's observed edge with a least-squares ellipse whose analytic center (Eq. 4) cancels the outward dilation caused by finite laser spot size. On the camera side, four ArUco markers on the same board give board pose, from which the four hole centers are derived. A closed-form Kabsch solution minimizes the 3D-3D distance (Eq. 5), and stacking $4N$ correspondences from multiple captures makes the joint optimization a single registration problem. The paper's evidence is five repeated calibrations on each of three LiDAR-camera pairs, with residuals below 6.5 mm and runtime under 0.7 s.","pith_inferences":["The angular-gap detector is a local geometric test, so in principle it should transfer to other depth sensors with irregular point spacing, provided enough points land on the hole edge; that is an extrapolation beyond the three LiDAR models tested.","The reported 6.5 mm residual is a registration residual on the calibration board, not a direct measurement of true extrinsic error; an independent ground-truth pose would separate the two.","Because one capture is so fast, the method could support periodic online re-calibration while a board is briefly in view, if the board's approximate location can be supplied automatically.","The ellipse-fitting correction for spot spread may also improve other circle- or hole-based calibration targets, since the dilation artifact is not specific to this board."],"forward_implications":["A calibration that takes under one second makes extrinsic re-calibration practical on production lines and in field deployment, not just in a dedicated off-line step.","The same board and code cover mechanical and solid-state LiDARs, so a mixed fleet can be calibrated with one procedure instead of one procedure per scan pattern.","Multi-scene joint calibration costs almost no extra time because captures are processed in parallel and the registration step is a single linear-algebra solve.","The 25-degree angular-gap rule is a purely geometric test, so new LiDAR models with unfamiliar scan patterns can be added without redesigning the detector."],"supporting_citations":[{"why":"Supplies the 3D structured target design and the baseline method that FAST-Calib compares against.","marker":"[13]"},{"why":"Documents the limitations of existing target-based calibration that FAST-Calib addresses.","marker":"[14]"},{"why":"Provides the ArUco marker detection used to derive the camera-frame board pose.","marker":"[22]"},{"why":"Provides the direct least-squares ellipse fitting used to compensate spot-spread dilation.","marker":"[23]"}],"fun_headline_variants":["One-second LiDAR-camera calibration, sub-6.5mm error","Calibrate LiDAR-camera in <1s with <6.5mm error","LiDAR-camera extrinsics in <1s, <6.5mm error","One-shot target-based LiDAR-camera calibration, <6.5mm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The calibration is only as good as the assumption that each fitted ellipse's center in the LiDAR point cloud coincides with the physical hole center, even when sparse beams and spot spread distort the observed hole boundary.","fun_headline_variants_meta":{"raw":{"variants":["One-second LiDAR-camera calibration, sub-6.5mm error","Calibrate LiDAR-camera in <1s with <6.5mm error","LiDAR-camera extrinsics in <1s, <6.5mm error","One-shot target-based LiDAR-camera calibration, <6.5mm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001621,"raw_usage":{"total_tokens":6436,"prompt_tokens":916,"completion_tokens":5520,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":5431}},"tokens_in":532,"tokens_out":5520,"duration_ms":42506,"temperature":1.0,"reasoning_tokens":5431,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:53:51.599670+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a board with machined hole centers at known coordinates, with the LiDAR-camera transform independently measured (e.g., by a motion-capture reference), and vary the board's incidence angle. If the estimated extrinsics drift by more than the reported 6.5 mm residual when spot dilation is large or holes are viewed near tangentially, the assumption that fitted ellipse centers equal true hole centers is falsified.","supporting_citations":[{"cited_title":"Automatic extrinsic calibration method for lidar and camera sensor setups,","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D structured target design and the baseline method that FAST-Calib compares against."},{"cited_title":"Joint camera intrinsic and lidar-camera extrinsic calibration,","cited_arxiv_id":null,"evidence_quote":"Documents the limitations of existing target-based calibration that FAST-Calib addresses."}],"review_version":1}