{"id":"4a5f27f1-17a3-4eb0-8244-53a08666233a","arxiv_id":"2501.11088","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Multi-LiCa calibrates multiple LiDARs automatically from static scenes, with no targets or initial pose guess, using feature matching plus iterative scan merging to bridge non-overlapping fields of view.","lead":"A new open-source pipeline calibrates several car-mounted LiDAR sensors against each other without calibration targets, vehicle motion, or a starting guess. It matches point-cloud features first, then refines alignment, and can handle sensors whose views do not directly overlap.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fitness-score gate in Algorithm 1 is unvalidated: a wrong but high-fitness merge can cascade through the overlap chain, so the no-initial-guess and partial-overlap claims are not yet fully established.","rationale":"The reader's verdict is CONDITIONAL, and my concern supports keeping that verdict rather than moving to accept or reject. The most load-bearing element of the central claim is not merely that the FOV graph must be connected; it is that the algorithm can reliably know when a pairwise alignment is good enough to merge. The paper's only gate is the GICP fitness threshold of 0.2. This is a genuine soft spot because the merging strategy is specifically designed for setups where some LiDARs have no direct overlap with the target, so errors in early merges are inherited by later ones. A wrong but high-fitness merge would silently corrupt every downstream calibration, and a correct but low-fitness alignment would stop the algorithm. Neither failure mode is analyzed, and the threshold was tuned on the same datasets used for evaluation, so the generalizability claim rests on a hyperparameter that has not been shown to transfer. The reader identified this as a secondary fragile premise; I elevate it to the primary concern because the cascading merge is the paper's main algorithmic contribution. The concrete threshold-sweep and fitness-separation test directly probes whether the gate is reliable. If the test passes, the concern is resolved and the method is more credible. If it fails, the paper would need a more robust acceptance criterion or explicit reporting of fitness margins to support the claim that it handles disconnected-FOV setups without initial guesses. Since this is an empirical validation gap rather than a demonstrated internal inconsistency, conditional acceptance remains the appropriate verdict, with the sweep and sensitivity analysis as required additions.","tokens_in":9218,"tokens_out":4023,"duration_ms":43602,"concrete_test":"Using the released code, run Algorithm 1 on the 20 evaluation scenes with the fitness threshold swept over {0.05, 0.10, 0.20, 0.30, 0.40}. Record calibration RMSE and the number of successfully calibrated LiDARs per threshold. If results are stable over a wide range, the threshold is not a critical linchpin. Additionally, for every pairwise GICP computation, compute the fitness from the ground-truth initialization and from the coarse-alignment initialization, and plot the two distributions. If 0.2 does not separate correct from incorrect alignments, the merge gate can admit wrong transforms and the cascading-overlap claim is not yet demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is the iterative cloud-merging strategy (Algorithm 1) that calibrates LiDARs with only indirect FOV overlap. Every merge decision and the termination condition depend on a single binary test: the Open3D GICP fitness score must exceed 0.2 (Section III.B). This fitness is the ratio of inlier correspondences to target points; it is a coverage statistic, not a correctness certificate. A GICP run that converges to a local minimum with plausible correspondences can score above 0.2 even when the transform is wrong, especially in self-similar or sparse geometry. Because the accepted source cloud is then merged into the growing target cloud, such an error becomes part of the reference for all subsequently calibrated LiDARs, so a single false positive propagates along the whole A->B->C->T chain. Conversely, a correct alignment for a sensor pair with small overlap may score below 0.2 and cause premature termination, leaving LiDARs uncalibrated. The paper only states 'We found a score of greater 0.2 sufficient' and provides no distribution of fitness for correct versus incorrect alignments, no sensitivity analysis around the threshold, and no report of the actual fitness margins on the accepted pairs. Since Section III.D says parameters were optimized on the same evaluation datasets, the 0.2 threshold could be overfit to these scenes. The connected-FOV assumption is explicitly acknowledged and reasonable, but the threshold is the unexamined load-bearing gate that makes the merging strategy work at all.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Multi-LiCa, an automatic, motion- and targetless extrinsic calibration framework for multiple LiDARs. The method first computes coarse alignment using FPFH features on voxelized clouds and TEASER++, then refines with GICP. The key algorithmic contribution is the iterative merging strategy in Algorithm 1: pairwise GICP fitness scores are computed for all available source–target pairs, the pair with the highest fitness is accepted if its score exceeds 0.2, the source cloud is transformed and merged into the growing target cloud, and the process repeats until all LiDARs are calibrated or no pair passes the threshold. This allows calibration of sensors that lack direct FOV overlap by cascading overlaps. The approach is evaluated on ten scenes from the HeLiPR dataset and ten scenes from the EDGAR vehicle, with CROON as the only baseline, and errors are reported as RMSEs in translation and rotation relative to the available ground truth. The paper also includes a LiDAR-to-ground calibration method and states that the source code is available.","tokens_in":9666,"tokens_out":4591,"duration_ms":46445,"significance":"The manuscript addresses a practically relevant problem: calibrating multiple LiDARs with heterogeneous sensing characteristics and only partial or indirect FOV overlap, without manual initial guesses or external targets. The main strengths are the removal of manual initialization, the iterative cloud-merging strategy that extends calibration beyond direct overlap, the explicit statement of the connected-FOV assumption, and the availability of source code. If the reported results were established out-of-sample, the framework would be a useful open-source contribution to the multi-LiDAR calibration literature. However, the central success gate (the fitness threshold), the parameter tuning procedure, and the ground-truth references are not yet validated to the level needed for the quantitative claims; the evaluation is also limited to twenty scenes and a single baseline. These issues are load-bearing for the abstract's claims of generalizability and parity or superiority to existing methods.","major_comments":[{"comment":"The decision to accept a merge, and the termination of the whole algorithm, rest entirely on the Open3D GICP fitness score exceeding 0.2, but the manuscript provides no evidence that fitness separates correct from incorrect alignments. Fitness is the ratio of inlier correspondences to target points, i.e., a coverage statistic rather than a correctness certificate; a wrong local minimum with plausible correspondences can exceed 0.2, and because the accepted source cloud is merged into the reference for all subsequently calibrated LiDARs, a single false positive can cascade through the overlap chain. Conversely, a correct low-overlap pair may score below 0.2 and cause premature termination. Please report the distribution of fitness for correct versus incorrect alignments, the actual fitness margins on accepted pairs, and a sensitivity analysis of the threshold, preferably on data not used for tuning.","section":"Section III.B, Algorithm 1"},{"comment":"The parameters, including voxel size, FPFH radius, GICP settings, and the fitness threshold, were optimized on the same HeLiPR and EDGAR scenes that later appear in the results. The reported accuracy is therefore an in-sample fit rather than an out-of-sample prediction, which weakens the paper's generalizability claim. Please either tune on a separate subset and report test-set results, or show that the results are insensitive to parameter variations over a plausible range, including threshold values around 0.2.","section":"Section III.D and Section IV"},{"comment":"The two ground-truth references are not adequate for the reported absolute RMSEs. For HeLiPR, the paper states that there is a substantial deviation between the ground-truth pose and the data for the Aeva LiDAR and that the values \"have to be looked at in relative terms,\" yet the reported errors are absolute RMSEs against that same ground truth. For EDGAR, the ground truth comes from a manual alignment of 0.5 m cubes in RVIZ with no uncertainty estimate. Please report relative error comparisons, quantify the ground-truth uncertainty or its influence on the RMSEs, and state whether the conclusions change when the suspected Aeva ground-truth error is removed.","section":"Section IV.A and Section IV.B"},{"comment":"The evaluation compares only against CROON and covers ten scenes per dataset, with some CROON failures excluded from the plots (\"not visible in the plot\"). The abstract's claim of being \"on par or better in calibration accuracy than existing approaches\" cannot be supported by a single baseline and by plots that omit failures. Please include at least one additional recent open-source baseline, report all runs including failures, and provide per-scene and per-LiDAR numeric results rather than only aggregated RMSEs.","section":"Section IV"}],"minor_comments":[{"comment":"The sentence \"An iterative process was used to optimize both the voxel size and the radius size\" gives no details on the objective, the search range, or the stopping criterion; please provide this information or point to the exact configuration in the repository.","section":"Section III.A"},{"comment":"The pseudocode is ambiguous about the recomputation step: \"Remove all calibrations involving the source LiDAR\" is immediately followed by \"for each calibration in calibrations tmp involving the last calibrated source,\" and the conflict between these two operations needs clarification.","section":"Algorithm 1"},{"comment":"The results are presented only through plots of aggregated RMSEs; numeric tables with per-axis, per-LiDAR values, and confidence intervals would make the comparison reproducible and easier for readers to assess.","section":"Section IV"},{"comment":"The evaluation of existing open-source methods is described briefly and relies on the authors' ability to compile and run them; a short table summarizing which methods were attempted and why they were excluded would be helpful context for the choice of CROON as the only baseline.","section":"Section II.A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a conference-style contribution with a useful and available implementation, but the quantitative evaluation is currently too thin for the stated claims. The main concern is not the algorithmic idea itself but the unvalidated fitness gate and the in-sample parameter tuning; both are fixable with additional experiments and reanalysis. I would not reject on the current evidence, but I would require a revised evaluation that addresses the threshold sensitivity and the ground-truth caveats before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about Multi-LiCa. First, the core idea is genuinely new: instead of requiring every LiDAR to overlap the target, it iteratively merges clouds into a growing target, so sensors with only indirect overlap (A->B->C->T) can be calibrated with no initial guess. That is a real contribution, and the code is on GitHub. Second, the evaluation is not strong enough to back the headline claims of generality. Parameters were optimized on the same HeLiPR and EDGAR scenes that later appear in the results, so the reported accuracy is in-sample. Only one baseline (CROON) is compared, and it fails outright on EDGAR, which makes the comparison look lopsided. The HeLiPR ground truth is admittedly imperfect, and the EDGAR ground truth is a manual RVIZ alignment with no uncertainty, so the absolute error bars are shaky.\n\nThe biggest soft spot is the fitness threshold. Algorithm 1 decides that a merge is correct if the Open3D GICP fitness exceeds 0.2. That number is presented as 'found sufficient' with no sensitivity analysis, no distribution of fitness for correct versus incorrect alignments, and no report of actual margins. Because a merged cloud becomes the reference for downstream LiDARs, a single false positive at 0.2 propagates through the whole chain. The stress-test note is right: this is an unvalidated load-bearing gate. That said, the paper explicitly states the connected-FOV assumption and admits that more testing is needed, so this is not a hidden flaw. It is an under-supported claim.\n\nWhat the paper does well: the writing is clear, the method is plausible, the algorithm pseudocode is concrete, and the code release is a practical enabler. For someone integrating a multi-LiDAR rig with non-overlapping views, this is a useful tool and a fair starting point. The cascading-merge strategy is worth citing, and the LiDAR-to-ground calibration trick is a nice minor addition.\n\nMy recommendation: send it to peer review, but require out-of-sample evaluation, at least one baseline that actually works on the target setup, uncertainty quantification for ground truth, and a real sensitivity analysis of the fitness threshold. The paper deserves referee time because the idea and code are valuable, but the current evidence is conditional at best.","headline":"A genuinely useful cascading-overlap calibration idea and clean open-source implementation, but the empirical case is thin and the 0.2 fitness gate is the load-bearing assumption that needs real validation.","tokens_in":10060,"tokens_out":1445,"would_cite":true,"duration_ms":16137,"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":"Multi-LiCa calibrates multiple LiDARs automatically and targetless, with no initial pose guess, from partially overlapping views.","keywords":["LiDAR extrinsic calibration","multi-LiDAR calibration","targetless calibration","point cloud registration","FPFH","TEASER++","GICP","autonomous vehicles"],"falsifier":"Run Multi-LiCa on a rig where one sensor's field of view is completely disjoint from all others, so no overlap chain reaches the target, and observe that the algorithm terminates without calibrating that sensor; or, using two clouds with a known transform, plant a wrong alignment with a GICP fitness above 0.2 and check whether it is merged and propagated.","tokens_in":9024,"feed_emoji":"🚗","tokens_out":8797,"duration_ms":81028,"temperature":0.7,"pith_summary":"Multi-LiCa is a targetless, motionless extrinsic calibration method for multi-LiDAR systems that requires no initial transformation estimate, no extra sensors, and no calibration targets. It aligns each LiDAR pair through a coarse feature-based step using FPFH/TEASER++, then refines with GICP, and iteratively merges successfully aligned clouds into a growing target cloud so that sensors without a direct field-of-view overlap can be calibrated through a chain of overlaps. The authors evaluate on a four-LiDAR automotive setup (EDGAR) and a four-LiDAR heterogeneous dataset (HeLiPR), reporting that Multi-LiCa is on par with or better than the CROON baseline in accuracy while computing each calibration in about 3 seconds on a single CPU core. If this holds, the method removes a significant manual step from deploying multi-sensor autonomous vehicles and extends calibration to unusual sensor layouts where no single LiDAR sees all others.","feed_headline":"No initial guess: Multi-LiCa calibrates many LiDARs automatically","feed_subtitle":"The two-stage pipeline chains overlapping views, calibrating sensors that never share a field of view.","key_machinery":"The load-bearing mechanism is a two-stage registration pipeline followed by a greedy merging loop. FPFH descriptors computed on voxelized clouds are matched by TEASER++ to produce a coarse transformation without an initial guess; GICP then refines each source–target pair. The GICP fitness score, defined as the number of inlier correspondences divided by the total points in the target cloud, decides which alignment is accepted, using a threshold of 0.2. Accepted source clouds are transformed and merged into the growing target cloud, and all calibrations involving the newly merged sensor are recomputed, which is what lets sensors with only indirect overlap to the target be calibrated.","core_discovery":"The central claim is that the full multi-LiDAR extrinsic calibration problem can be solved automatically under a connected-overlap assumption: every LiDAR must share a field of view, directly or transitively, with a chosen target LiDAR. The pipeline computes a coarse transformation between every pair of voxelized point clouds using FPFH features matched by TEASER++, obtaining a global alignment without any user-provided initial guess. Each pair is then fine-registered with GICP, and the most successful alignment, measured by a fitness score of at least 0.2, is merged into the growing target cloud; all remaining pairs involving the newly merged sensor are recomputed against the merged cloud. This greedy iterative merge allows a sensor like EDGAR's rear LiDAR, which has no direct overlap with any other single sensor in the worst case, to be calibrated through intermediate clouds. The authors claim that on the EDGAR and HeLiPR datasets this approach is on par with or more accurate than the existing CROON method, especially for setups with angled, non-overlapping LiDARs, and is about three times faster.","pith_inferences":["The greedy merge policy makes calibration order-dependent: the first accepted high-fitness pair biases the growing cloud, so a different initial pairing could yield a different final calibration; a global consistency check across pairwise transformations would make the result more repeatable.","The fixed fitness threshold of 0.2 could admit a wrong alignment in symmetric or feature-poor scenes, so a validation step that re-checks the final multi-cloud consistency or rejects merges whose relative pose disagrees with other pairwise estimates would add robustness.","The coarse stage's choice of FPFH/TEASER++ over RANSAC-based feature matching is motivated by robustness; comparing the coarse stage against learned registration methods would clarify the accuracy/runtime trade-off for future setups."],"forward_implications":["Multi-LiDAR rigs with partial or chain-connected field-of-view overlaps can be calibrated automatically, without targets, motion, or initial pose estimates, as long as every sensor has a direct or transitive overlap path to the target LiDAR.","The coarse-to-fine design with FPFH/TEASER++ and GICP removes the need for manual initialization, which the paper shows is critical for angled setups like EDGAR where existing methods cannot align the rear LiDAR.","Because the iterative merge recomputes alignments against growing clouds, the method scales to any number of sensors with a single parameter set, and the paper demonstrates this on two four-LiDAR setups with different sensor types and scan patterns.","The optional LiDAR-to-ground calibration step computes roll, pitch, and vertical offset to a base frame from a stationary, flat ground plane, further automating the sensor-to-vehicle calibration."],"supporting_citations":[{"why":"Supplies the Fast Point Feature Histogram descriptor used to compute coarse-alignment features on voxelized point clouds.","marker":"[35]"},{"why":"Provides TEASER++, the robust registration method that matches FPFH features into a coarse transformation without an initial guess.","marker":"[36]"},{"why":"Defines Generalized-ICP, the fine registration algorithm applied to every source–target pair.","marker":"[22]"},{"why":"The GICP implementation that supplies the fitness score and transformation refinement used in the iterative merging loop.","marker":"[37]"},{"why":"CROON, the existing multi-LiDAR calibration method used as the baseline on both evaluation datasets.","marker":"[10]"},{"why":"HeLiPR, the heterogeneous LiDAR dataset with ground-truth poses that Multi-LiCa is evaluated on.","marker":"[38]"}],"fun_headline_variants":["No targets, no initial guess: Multi-LiCa auto-calibrates LiDARs","Multi-LiCa chains overlapping views to calibrate every LiDAR","Automatic multi-LiDAR calibration without any manual input","Calibrate LiDARs without direct overlap using transitive views","Multi-LiCa: accurate, targetless calibration for any LiDAR setup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that every LiDAR's field of view overlaps with the target either directly or through a chain of pairwise overlaps, so that the iterative merging can eventually include every sensor; a LiDAR with no overlap to any already-calibrated cloud terminates the algorithm without being calibrated.","fun_headline_variants_meta":{"raw":{"variants":["No targets, no initial guess: Multi-LiCa auto-calibrates LiDARs","Multi-LiCa chains overlapping views to calibrate every LiDAR","Automatic multi-LiDAR calibration without any manual input","Calibrate LiDARs without direct overlap using transitive views","Multi-LiCa: accurate, targetless calibration for any LiDAR setup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000749,"raw_usage":{"total_tokens":3347,"prompt_tokens":965,"completion_tokens":2382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":2287}},"tokens_in":581,"tokens_out":2382,"duration_ms":20415,"temperature":1.0,"reasoning_tokens":2287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:38:45.151528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Multi-LiCa on a rig where one sensor's field of view is completely disjoint from all others, so no overlap chain reaches the target, and observe that the algorithm terminates without calibrating that sensor; or, using two clouds with a known transform, plant a wrong alignment with a GICP fitness above 0.2 and check whether it is merged and propagated.","supporting_citations":[{"cited_title":"Fast point feature histograms (fpfh) for 3d registration,","cited_arxiv_id":null,"evidence_quote":"Supplies the Fast Point Feature Histogram descriptor used to compute coarse-alignment features on voxelized point clouds."},{"cited_title":"TEASER: Fast and Certifiable Point Cloud Registration,","cited_arxiv_id":null,"evidence_quote":"Provides TEASER++, the robust registration method that matches FPFH features into a coarse transformation without an initial guess."},{"cited_title":"Generalized-icp,","cited_arxiv_id":null,"evidence_quote":"Defines Generalized-ICP, the fine registration algorithm applied to every source–target pair."},{"cited_title":"Generalized icp registration — open3d 0.17.0 documenta- tion,","cited_arxiv_id":null,"evidence_quote":"The GICP implementation that supplies the fitness score and transformation refinement used in the iterative merging loop."},{"cited_title":"CROON: Automatic Multi-LiDAR Calibration and Refinement Method in Road Scene,","cited_arxiv_id":null,"evidence_quote":"CROON, the existing multi-LiDAR calibration method used as the baseline on both evaluation datasets."},{"cited_title":"Helipr: Heterogeneous lidar dataset for inter-lidar place recognition under spatial and temporal variations,","cited_arxiv_id":null,"evidence_quote":"HeLiPR, the heterogeneous LiDAR dataset with ground-truth poses that Multi-LiCa is evaluated on."}],"review_version":1}