{"id":"7eb15bbd-154b-4cfa-9c9e-dff56862961a","arxiv_id":"2507.09176","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DLBAcalib calibrates non-overlapping LiDARs by building a map from one LiDAR with bundle adjustment and aligning the other LiDAR's frames to that map, achieving about 5 mm and 0.2 degree error in simulation.","lead":"This paper introduces a LiDAR calibration method that works when two sensors do not see the same area at the same time, by aligning one sensor's scans to a map built by the other. It claims centimeter-level accuracy and tolerance to rough initial guesses, which could make multi-LiDAR setups easier to deploy in vehicles and robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reference map is fixed during extrinsic optimization, so map drift is absorbed into T_AB; the 5 mm/0.2° claim is not tested against this bias and the 'joint LBA' description is not reflected in Eq. (7).","rationale":"The reader's weakest assumption identifies the accuracy of C_map_A as the load-bearing condition, and I agree that this is the central risk. My stress-test makes that concern more precise: the formulation in Eq. (7) fixes C_map_A during extrinsic optimization, so map errors are not corrected by any joint refinement, and no global consistency mechanism is present. The paper's own Fig. 6 tests localization degradation but does not report the corresponding T_AB errors, leaving a quantitative gap. Algorithm 1 also introduces an internal inconsistency in how T_AB is updated, which weakens the reproducibility of the claimed results. I do not see grounds to reject the paper outright: the simulation results are plausible, the code is open source, and the basic frame-to-map alignment idea can work under accurate odometry. However, the central claim of 5 mm/0.2° and robustness to map error needs the specific verification above. Since the reader's verdict was already CONDITIONAL with moderate confidence, my concern does not change the verdict; it refines the condition that should be attached. I therefore mark agreement as partial and keep the verdict unchanged.","tokens_in":12155,"tokens_out":4690,"duration_ms":63141,"concrete_test":"Open the released code and check whether C_map_A is ever re-optimized after aligning LiDAR B frames; if it is not, state that 'joint LBA' is not implemented. Then run the pipeline on the same CARLA sequences twice: once with ground-truth LiDAR A poses and once with FastLIO2+LBA poses used to build C_map_A. Compare the estimated T_AB in both runs. Additionally, inject a known global deformation into C_map_A (e.g., a 1 cm/m shear or a constant yaw drift) and observe the shift in T_AB. If the extrinsic estimate changes by more than the reported 5 mm/0.2°, the map-bias channel is real and the headline accuracy needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that C_map_A, built once from FastLIO2 plus sliding-window LBA in Eq. (3), is an unbiased reference. However, Eq. (7) optimizes only T_AB while treating C_map_A as fixed; the abstract's 'joint LBA optimization' is therefore not implemented as a joint refinement of the map and the extrinsics. Sliding-window LBA only enforces local consistency within each window and does not provide global loop closure, so any residual drift or deformation in C_map_A is not corrected by the calibration stage. Because the point-to-plane residual in Eq. (6) measures distance to a potentially warped plane, map error is indistinguishable from extrinsic error and will bias T_AB. The paper's trajectory-degradation experiments (Fig. 6) do not report the resulting T_AB error conditioned on map accuracy, so the headline accuracy of 5 mm and 0.2° is not shown to be robust to the dominant error channel. Algorithm 1 compounds the concern: it computes per-frame local alignments and averages them, and line 13 applies Tavg * Tguess, which is inconsistent with the iterative update in Eq. (9); this makes the actual estimated quantity ambiguous. The weakest link is thus the assumption that residuals in Eq. (6) reflect only extrinsic error, with no demonstrated mechanism to reject or absorb reference-map bias.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DLBAcalib, a targetless extrinsic calibration method for dual LiDARs with non-overlapping fields of view. LiDAR A scans are deskewed, processed by FastLIO2, and refined by sliding-window LiDAR bundle adjustment to construct a reference map C_map_A. Adaptive voxelization extracts planar features with confidence weights. LiDAR B frames are then aligned to this map by minimizing weighted point-to-plane residuals over the extrinsic transform T_AB via Levenberg-Marquardt. Experiments in the CARLA simulator across five configurations and in three real-world backpack-mounted scenes report an average translation error of 5 mm and rotation error of 0.2 deg, with tolerance to initial errors up to 0.4 m and 30 deg, compared against MLCC and PCDcalib.","tokens_in":12479,"tokens_out":6011,"duration_ms":77343,"significance":"If the claimed performance holds, the method is practically valuable: it removes the need for overlapping fields of view and precise initialization, uses only LiDAR data plus IMU, and is released as open-source code. The simulation study uses ground-truth extrinsics and multiple sensor configurations, and the comparison against two existing calibration methods is useful. However, the central 'joint LBA' claim is broader than the implemented optimization, and Algorithm 1 as written is internally inconsistent. These issues must be resolved before the accuracy and robustness claims can be accepted. The open-source release and reproducible simulation pipeline are strengths that make the required revisions feasible.","major_comments":[{"comment":"The abstract and the contribution list state that extrinsic calibration is formulated as a joint LBA optimization and that the method 'jointly refines intra-source and inter-source point cloud frames'. The objective in Eq. (7), however, depends only on T_AB; C_map_A is built once in Eq. (3) and held fixed during extrinsic optimization. The sliding-window LBA in Section III-D optimizes only LiDAR A's poses, not the extrinsics. As written, the method is a two-stage pipeline: LBA-refined mapping followed by frame-to-map extrinsic optimization. This mismatch with the central claim is load-bearing. Please either revise the claims to describe a two-stage method, or extend the optimization to jointly refine the map poses and T_AB and report the corresponding objective and Jacobians.","section":"Abstract and Section III-F, Eq. (7)"},{"comment":"Algorithm 1 does not implement the optimization described by Eqs. (6)-(9). It runs a separate BA per frame starting from Tprev (line 4), averages the resulting transforms (line 6), and returns Tavg * Tguess (line 13). If each T_local is already the result of the iterative update in Eq. (9) initialized at Tprev, then line 13 applies the initial guess a second time. If T_local is instead an incremental correction, that is never stated. The averaging of SE(3) transforms is also undefined. Moreover, optimizing each frame independently from the same initial value and averaging the results is not equivalent to minimizing the global objective in Eq. (7) over all frames. The actual estimated quantity is therefore ambiguous. Please specify the exact per-frame optimization, the averaging operation, the final composition, and how the convergence test relates to Eq. (9).","section":"Section III-F, Algorithm 1 and Eq. (9)"},{"comment":"The headline accuracy depends on C_map_A being an unbiased geometric reference. The sliding-window LBA in Section III-D enforces local consistency within each window but does not provide global loop closure, so residual drift or warping in C_map_A is fixed during the extrinsic stage. Since the residual in Eq. (6) measures the distance to planes extracted from this map, map error is indistinguishable from extrinsic error and will bias T_AB. Figure 6 reports trajectory degradation but does not report the resulting T_AB error conditioned on map accuracy. Please provide quantitative results showing T_AB error under controlled map perturbation or on trajectories without loop closure, or otherwise show that the 5 mm / 0.2 deg result is robust to this dominant error channel.","section":"Section III-D and Section III-F, Eq. (6)"},{"comment":"Tables III and IV report only mean translation and rotation errors over trials. The claims of 'consistent performance' and 'initial error tolerance up to 0.4 m and 30 deg' require standard deviations, medians, maxima, and failure rates over the 50 runs. Without these, the reader cannot assess whether the mean is driven by a few good runs or whether all runs converge. Please report the full error distributions and also provide the exact values used for the thresholds and hyperparameters: window size w, step d, convergence threshold delta, maximum iterations N, LM damping mu, and the voxel parameters tau_eta, tau_theta, tau_d, and gamma.","section":"Tables III and IV"},{"comment":"The real-world evaluation reports no ground-truth extrinsics and only qualitative visual comparisons in Figs. 7 and 8. This cannot support the claim that DLBAcalib 'consistently outperformed' MLCC and PCDcalib in real scenes. Please provide quantitative metrics, such as point-to-plane residuals on held-out scans, plane-fit consistency, or independent motion-based ground truth, and report per-scene numbers for each method.","section":"Section III-K, Real-World Experiment Results"},{"comment":"The phrase 'outlier-resistant parameter estimation' is not supported by Eqs. (5) and (7) as written. The confidence weight omega_v depends on point count, eigenvalue spread, and planarity, but not on residual magnitude. There is no robust loss function, no residual-based reweighting, and no explicit outlier rejection step described. Scenario 2 is only shown qualitatively. Please specify the mechanism that makes the estimator robust to outliers and evaluate it quantitatively, for example by injecting outliers into simulated scans and reporting T_AB error.","section":"Section III-E and Eq. (5)"}],"minor_comments":[{"comment":"The sentence beginning 'To intuitively explain how precise calibration is achieved without overlapping FoVs,one forward-facing (LiDAR A) and one rear-facing (LiDAR B), with non-overlapping FoVs.' is grammatically incomplete and should be rewritten.","section":"Section III-B"},{"comment":"The averaging expression uses 't=n to t=0' while the loop in line 3 runs 't=1 to n'; use consistent index ranges and define the SE(3) averaging operation explicitly.","section":"Algorithm 1, line 6"},{"comment":"The Jacobian in Eq. (8) is consistent with the right-perturbation update in Eq. (9), but the sign and the use of (p_i)^wedge versus (R p_i)^wedge are not derived; a one-sentence derivation would prevent confusion.","section":"Section III-F, Eq. (8)"},{"comment":"Figure 6 does not define how 'trajectory error' is measured or how the degraded trajectories with 5 cm and 10 cm error were generated; please add this information to the caption or text.","section":"Figure 6"},{"comment":"The adaptive voxelization thresholds tau_eta, tau_theta, tau_d, gamma, and the maximum recursion depth are not given; provide these values for reproducibility, especially since the paper claims no manual parameter tuning.","section":"Section III-E"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and the simulation framework is appropriate, but the manuscript is not ready in its current form. The two barriers are the contradiction between the claimed joint LBA and the implemented two-stage optimization, and the internal inconsistency of Algorithm 1 with Eq. (9). Both are fixable with a careful rewrite and additional experiments. I would not reject, because the open-source code and ground-truth simulation setup give the authors a clear path to address the concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on DLBAcalib. The practical idea is sound: use sliding-window LBA to build a clean reference map from one LiDAR, then calibrate the second by aligning its frames to that map with point-to-plane constraints. That is a sensible way to handle non-overlapping FoVs, and the demonstrated tolerance to 0.4 m and 30° initial errors is a meaningful practical improvement over what MLCC and PCDcalib advertise. The code is open source and the simulation experiments include ground truth, which gives the numbers some teeth.\n\nThe soft spots are real, in ascending order of concern.\n\nThird, the title and abstract claim 'joint LBA optimization' of the map and extrinsics, but Eq. (7) only optimizes T_AB while C_map_A is fixed. The LBA is used only to build the map, not during calibration. That's a terminology problem more than a technical one, but it conflates the two stages.\n\nSecond, Algorithm 1 ends with 'T_AB <- Tavg * Tguess'. If Tavg is the averaged result from optimizing with Tprev initialized to Tguess, that multiplication double-counts the initial guess. This looks like a typo, but it means the algorithm's output is ambiguous as written. The authors should clarify whether Tavg is the new estimate or the incremental correction.\n\nFirst, the accuracy claim. The 5 mm and 0.2° numbers are from simulation only. Real-world experiments have no ground truth. More importantly, the paper claims robustness to front-end trajectory error, but it never reports the resulting extrinsic error as a function of trajectory degradation. A warped reference map will bias T_AB, and the point-to-plane residuals in Eq. (6) cannot distinguish map error from extrinsic error. The stress-test note is right: the current experiments don't rule out that the method's accuracy is limited by map drift.\n\nOne minor point: the planarity index in Eq. (4) uses an unusual ordering of eigenvalues; the authors should specify whether eigenvalues are sorted ascending or descending, otherwise the threshold η < 0.1 is ambiguous.\n\nBottom line: this deserves a serious referee. The flaws are fixable with a clearer algorithm description, error bars, and a map-bias ablation. The target audience is robotics practitioners who need cheap multi-LiDAR setups without overlapping views. I would send it to review with a request for major revision.","headline":"Promising non-overlapping LiDAR calibration, but the 'joint LBA' claim is overstated and the accuracy numbers need stronger support.","tokens_in":12991,"tokens_out":6493,"would_cite":false,"duration_ms":70653,"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":"Dual LiDARs without overlapping views can be calibrated to 5 mm accuracy using bundle adjustment on a reference map built from one sensor's scans.","keywords":["LiDAR extrinsic calibration","non-overlapping field of view","targetless calibration","LiDAR bundle adjustment","adaptive voxelization","point-to-plane registration","robust optimization","multi-LiDAR system"],"falsifier":"Run the calibration in an environment with weak planar structure (e.g., an open field with bushes) or with several moving vehicles, and measure whether the translation and rotation errors stay near 5 mm and 0.2 degrees; if the errors grow proportionally to the map's planarity index or outlier ratio, the assumption that point-to-plane residuals reflect extrinsic error rather than map error is violated.","tokens_in":11994,"feed_emoji":"📡","tokens_out":2265,"duration_ms":27739,"temperature":0.7,"pith_summary":"This paper claims that two LiDARs with non-overlapping fields of view can be extrinsically calibrated without artificial targets, precise initial guesses, or manual tuning. The key move is to first build a geometrically accurate reference map from one LiDAR using sliding-window LiDAR bundle adjustment, then align the other LiDAR's frames to that map with point-to-plane residuals. The authors report average translational errors of 5 mm and rotational errors of 0.2 degrees, while tolerating initial extrinsic errors up to 0.4 m and 30 degrees. If the claim holds, multi-LiDAR systems can be deployed in arbitrary configurations with only coarse installation knowledge, which would simplify sensor setup for autonomous driving and 3D reconstruction.","feed_headline":"Twin LiDAR calibration reaches 5 mm with no overlap","feed_subtitle":"A bundle-adjusted reference map from one sensor lets the second align without targets or precise starting guesses.","key_machinery":"The load-bearing mechanism is a dual LiDAR bundle adjustment (LBA) pipeline: sliding-window LBA (using the BALM-style optimization) refines the target LiDAR's pose trajectory to build a clean reference map, adaptive voxelization extracts planar features with confidence weights, and a Levenberg-Marquardt optimizer minimizes the weighted point-to-plane residuals parameterized through se(3) Lie algebra. This frame-to-map alignment substitutes for direct overlap, while the adaptive weighting suppresses outliers and the LBA step reduces cumulative odometry drift.","core_discovery":"The central claim is that extrinsic calibration for non-overlapping LiDARs can be recast as a frame-to-map alignment problem in which one LiDAR's accumulated point cloud serves as a fixed reference, eliminating the need for simultaneous field-of-view overlap. The paper argues that by optimizing the target LiDAR's poses through sliding-window LBA, the reference map becomes accurate enough that point-to-plane residuals between the other LiDAR's frames and the map mostly capture extrinsic error rather than map distortion. A robust iterative refinement with adaptive voxel weighting then yields convergence from coarse initial estimates, achieving a translational error of 5 mm and a rotational error of 0.2 degrees in dual-LiDAR setups.","pith_inferences":["The accuracy ceiling is set by the quality of the LBA-optimized reference map; if the scene lacks strong planar structure or contains many moving objects, the reported 5 mm/0.2° figures would likely degrade, so the method is best suited to structured static environments.","A natural testable extension is to use the calibrated extrinsic to check map consistency between the two LiDARs (e.g., by re-aligning a map built from LiDAR B back to LiDAR A's map), which would expose any residual bias that the point-to-plane residual alone may not reveal.","The sliding-window LBA step implicitly assumes the IMU-based odometry poses are good enough to bootstrap window optimization; in GNSS-denied or highly dynamic trajectories, the reference map could warp, introducing a proportional bias into the extrinsic estimate.","Because the residual is purely geometric, the method could be extended to calibrate a LiDAR against a camera's structure-from-motion map or against an offline 3D model, provided corresponding planar features exist."],"forward_implications":["Multi-LiDAR vehicles could be calibrated in minutes from a short drive, without placing targets or measuring mounting offsets by hand.","The method removes the constraint that all LiDARs must observe a common region, permitting flexible sensor placement such as one forward-facing and one rear-facing unit.","Accuracies around 5 mm and 0.2 degrees would make the calibrated extrinsics usable directly in downstream mapping and localization, where sub-centimeter errors are typically required.","Since the reference map is built from the target LiDAR alone, the approach can be extended to any number of secondary LiDARs as long as each observes the same environment as the map at some time.","The tolerance to initial errors of 0.4 m and 30 degrees means a rough manual measurement of the sensor mount is sufficient to start the calibration."],"supporting_citations":[{"why":"FastLIO2 provides the front-end LiDAR-inertial odometry poses that initialize the sliding-window LBA and the frame-to-map alignment.","marker":"[22]"},{"why":"BALM supplies the efficient bundle adjustment method used to optimize the target LiDAR's pose trajectory within each sliding window.","marker":"[23]"},{"why":"The adaptive voxelization strategy is adopted to extract planar features from the reference map, which the point-to-plane residuals depend on.","marker":"[24]"},{"why":"CARLA is used to generate simulated LiDAR data with ground-truth extrinsics and configurable noise, providing the primary quantitative evaluation.","marker":"[25]"},{"why":"MLCC is a target-free bundle-adjustment baseline that the paper compares against, especially for non-overlapping configurations.","marker":"[21]"},{"why":"PCDcalib (ICP-based frame-to-map calibration) is the other main baseline, representing the prior approach that does not refine the reference map.","marker":"[20]"}],"fun_headline_variants":["No-overlap LiDAR calibration reaches 5 mm","Targetless LiDAR calibration hits 5 mm precision","Sliding-window LBA calibrates LiDARs to 5 mm","Robust LiDAR calibration: 5 mm, no targets, no overlap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the bundle-adjusted reference map built from one LiDAR is accurate enough that misalignment between the second LiDAR's frames and the map is mostly due to extrinsic error, not map warping or dynamic objects.","fun_headline_variants_meta":{"raw":{"variants":["No-overlap LiDAR calibration reaches 5 mm","Targetless LiDAR calibration hits 5 mm precision","Sliding-window LBA calibrates LiDARs to 5 mm","Robust LiDAR calibration: 5 mm, no targets, no overlap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001076,"raw_usage":{"total_tokens":4499,"prompt_tokens":936,"completion_tokens":3563,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":3490}},"tokens_in":552,"tokens_out":3563,"duration_ms":28734,"temperature":1.0,"reasoning_tokens":3490,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:01:19.153002+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the calibration in an environment with weak planar structure (e.g., an open field with bushes) or with several moving vehicles, and measure whether the translation and rotation errors stay near 5 mm and 0.2 degrees; if the errors grow proportionally to the map's planarity index or outlier ratio, the assumption that point-to-plane residuals reflect extrinsic error rather than map error is violated.","supporting_citations":[{"cited_title":"Fast-LIO2: Fast direct LiDAR-inertial odometry,","cited_arxiv_id":null,"evidence_quote":"FastLIO2 provides the front-end LiDAR-inertial odometry poses that initialize the sliding-window LBA and the frame-to-map alignment."},{"cited_title":"Efficient bundle adjustment for LiDAR SLAM with BALM,","cited_arxiv_id":null,"evidence_quote":"BALM supplies the efficient bundle adjustment method used to optimize the target LiDAR's pose trajectory within each sliding window."},{"cited_title":"Efficient adaptive voxelization for LiDAR point cloud processing,","cited_arxiv_id":null,"evidence_quote":"The adaptive voxelization strategy is adopted to extract planar features from the reference map, which the point-to-plane residuals depend on."},{"cited_title":"CARLA: An open urban driving simulator,","cited_arxiv_id":null,"evidence_quote":"CARLA is used to generate simulated LiDAR data with ground-truth extrinsics and configurable noise, providing the primary quantitative evaluation."},{"cited_title":"Targetless extrinsic calibration for multi-LiDAR systems using adaptive voxelization and bundle adjust- ment,","cited_arxiv_id":null,"evidence_quote":"MLCC is a target-free bundle-adjustment baseline that the paper compares against, especially for non-overlapping configurations."},{"cited_title":"High-precision extrinsic calibration of multiple LiDARs using ICP,","cited_arxiv_id":null,"evidence_quote":"PCDcalib (ICP-based frame-to-map calibration) is the other main baseline, representing the prior approach that does not refine the reference map."}],"review_version":1}