{"id":"355e019b-58ce-45b6-81a8-afec03b93462","arxiv_id":"2507.23595","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MamV2XCalib fuses multi-frame vehicle LiDAR projections with roadside camera images, using 4D correlation volumes and Mamba temporal fusion to regress the camera's rotation error.","lead":"This paper introduces MamV2XCalib, a target-less method for calibrating roadside cameras using LiDAR data from passing vehicles, without reference objects or manual intervention. It reports accurate rotation estimates on two real-world V2X datasets and is released as open-source code.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never specifies how moving-vehicle LiDAR is transformed into the fixed roadside camera frame; Eqs. 1–3 treat TLC as a constant, which is geometrically inconsistent unless vehicle pose is applied per frame.","rationale":"I read the paper as a systems/calibration paper whose main contribution is the application of an iterative correlation-volume plus Mamba architecture to V2X roadside camera calibration. The network components are reasonably supported by ablations (Tables 2 and 4), and the code is promised, which counts as independent support if it is provided. The weakest place is not the network but the geometric setup: the paper's equations and text do not specify how each moving-vehicle LiDAR frame is related to the fixed roadside camera. The reader's weakest_assumption identifies exactly this issue, and I agree it is load-bearing because every experiment, loss term, and reported error depends on the projection being geometrically coherent. The concern is resolvable: if the released code applies per-frame vehicle pose compensation before projection, the method is plausible and the conditional verdict stands; if not, the numerical results are suspect. Since the reader already made the verdict CONDITIONAL on this issue, my stress-test does not move the verdict; it sharpens the required verification. No ad hominem or consensus-based objection is intended; this is purely an internal consistency and reproducibility question about the coordinate frames.","tokens_in":12190,"tokens_out":6115,"duration_ms":70898,"concrete_test":"Inspect the released code path that generates the input depth maps and verify whether each LiDAR scan is first transformed by the vehicle pose into a common world frame before applying the erroneous camera extrinsic. Analytically, derive Eq. 1 with a time index: for a static world point X, the correct projection is p_cam(t) = K · ΔT · T_cam_world · X, while the paper's constant-TLC form gives K · ΔT · TLC · p_lidar(t). Show whether these agree for all t; if they do not, the central calibration error numbers are not measuring the claimed quantity. If pose compensation is present in code, request a revised Sec. 3.1 that states this explicitly and updates Eqs. 1–3 and Eq. 11 accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is targetless V2X camera calibration with mean rotation error 0.6313° and std 0.3211° on V2X-Seq. That claim presupposes that each vehicle LiDAR scan can be projected into the fixed roadside camera using a single erroneous extrinsic matrix. The load-bearing fragility is the frame bookkeeping. Eq. 1 states A is the vehicle LiDAR coordinate system and R,t transform from the onboard LiDAR to the camera; Eq. 2 writes Tinit = ΔT·TLC with TLC described as the transformation from 'the current LiDAR coordinate system' to the camera. But the LiDAR is moving while the camera is fixed, so TLC is time-dependent: TLC(t) = T_cam_world · T_world_veh(t) · T_veh_lidar. The paper does not state whether each point cloud is first brought into a common world frame using vehicle pose. Sec. 3.1 only says data is 'aligned to a unified coordinate system' and then projected with Tinit, and Sec. 3.4 acknowledges the vehicle is in motion but does not explain how this is handled. If a constant Tinit is applied directly to raw scans, the depth maps are not valid projections of the scene; the calibration flow then encodes vehicle ego-motion confounded with camera rotation error, so the reported errors do not measure what they claim. If the implementation does apply pose compensation, then Eqs. 1–3 and Eq. 11 are formally wrong and the revision must make the per-frame transform explicit. This is not a stylistic issue: it determines whether the central quantitative claim is meaningful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MamV2XCalib, a targetless V2X-based method for calibrating the rotation of a fixed roadside camera using vehicle-mounted LiDAR and roadside camera images. The method projects vehicle LiDAR point clouds into the camera view using an initially erroneous extrinsic transform, computes a pixel-level 'calibration flow' via multi-scale 4D correlation volumes and iterative GRU updates, and then uses a Mamba-based spatiotemporal fusion module to regress the camera rotation error from a sequence of such flow fields. The authors report mean rotation errors of 0.6313 degrees (std 0.3211) on V2X-Seq for initial deviations up to ±20 degrees, and 0.267 degrees on TUMTraf-V2X for ±5 degrees, outperforming LCCNet and a zero-training baseline. The paper claims to be the first targetless V2X-based roadside camera calibration method and provides code.","tokens_in":103,"tokens_out":4199,"duration_ms":99385,"significance":"If the geometric consistency issue described below is resolved, the paper makes a practical contribution: it offers a deployable, targetless calibration approach that exploits existing vehicle infrastructure perception data, with code, real-world evaluations on two datasets, and thorough ablations showing the value of temporal fusion and iterative refinement. The Mamba-based temporal aggregation is a technically sensible design for handling sparse, time-varying depth maps in V2X settings. However, the central quantitative claims currently rest on an underspecified and potentially incorrect coordinate-frame bookkeeping, which must be fixed before the results can be considered meaningful.","major_comments":[{"comment":"The manuscript defines the source coordinate system A as the vehicle LiDAR coordinate system and states that TLC is the transformation from the current LiDAR coordinate system to the camera coordinate system. Since the LiDAR is on a moving vehicle while the camera is fixed, TLC is necessarily time-dependent: TLC(t) = T_cam_world * T_world_veh(t) * T_veh_lidar. The paper never specifies whether each point cloud is first transformed into a common world frame using the vehicle pose before projection with Tinit. The sentence in Sec. 3.1 that data is 'aligned to a unified coordinate system' is too vague to establish the bookkeeping. If a constant Tinit is applied directly to raw scans, the resulting depth maps are not valid projections of the static scene; the calibration flow would then confound vehicle ego-motion with camera rotation error, invalidating the reported mean errors and standard deviations. If pose compensation is in fact applied, Eqs. (1)-(3) and the surrounding text must be rewritten to make the per-frame transform explicit. This is the load-bearing geometric assumption of the entire method.","section":"Sec. 3.1, Eqs. (1)-(3)"},{"comment":"The point-cloud loss Lp = (1/N) sum_i || T_LC^{-1} T_pred^{-1} T_init P_i - P_i ||^2 uses a single ground-truth LiDAR-camera extrinsic matrix T_LC for all points P_i. If the vehicle moves during data collection, this loss is only correct if every P_i has been transformed into a common world frame before the loss is computed. The paper does not state how P_i are obtained or accumulated, and the notation suggests that all points are in the same LiDAR frame at a single time instant. This is the same frame-bookkeeping issue as in Sec. 3.1, but it directly affects the training objective: if the loss is computed in the raw vehicle LiDAR frame, it does not measure what the paper claims. The authors must clarify the coordinate frame of P_i in Eq. (11) and provide the corresponding per-frame transforms.","section":"Sec. 3.5, Eq. (11)"},{"comment":"The network is trained by sampling random perturbations Delta_T from a given range and applying them to TLC to generate Tinit, and then evaluated on the same perturbation distribution. This is a standard supervised protocol for learning-based calibration, but the paper's robustness claims would be stronger if the authors reported performance on perturbation ranges not seen during training (e.g., train on ±10 degrees, test on ±20 degrees, or vice versa). Without such experiments, the reported standard deviations only demonstrate within-distribution stability, not generalization to larger or differently distributed initial errors. The distance filtering introduced in Sec. 4.3 also means that the reported numbers apply only to the filtered subset; reporting unfiltered results or a distance sweep would give a more complete picture.","section":"Sec. 4.2 and Sec. 4.3"}],"minor_comments":[{"comment":"The text refers to 'Appendix 4.3' but the paper does not contain a numbered appendix; please fix the cross-reference.","section":"Sec. 4.3"},{"comment":"The rows for Calib-anything are entirely dashes; the caption says it 'fails,' but the table would be clearer if the failure were briefly described (e.g., 'diverges and produces unusable results') in the caption or a footnote.","section":"Table 1"},{"comment":"The row '3D feature matching [21]' uses dashes for the temporal, iteration, and correlation-volume columns, which is ambiguous; clarify that this row replaces the entire matching module with LCCNet's 3D matching layer.","section":"Table 2"},{"comment":"The notation T0, T1, ... Tn in Eq. (13) is not precisely defined; state whether these are the outputs of successive network passes, and how the chain is initialized and terminated.","section":"Sec. 3.6, Eq. (13)"},{"comment":"The image encoder uses a 'pre-trained ResNet-18'; please specify the pre-training dataset (e.g., ImageNet) and whether the depth encoder is trained from scratch, as this affects reproducibility.","section":"Sec. 3.2"},{"comment":"The argument for ignoring translation noise is supported by Fig. 5, but the corresponding experiment in the appendix is only mentioned in passing; include the quantitative result or a reference to the appendix section where it appears.","section":"Sec. 4.5"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the coordinate-frame bookkeeping. The paper must clarify whether vehicle poses are used to transform LiDAR scans into a common world frame before projection. If they are, the method section needs a careful rewrite; if they are not, the reported results are likely confounded by ego-motion and the central claims are not supported. Please also ask for cross-range generalization experiments to substantiate the robustness claim. The paper is otherwise a reasonable engineering contribution with code and real-data experiments, but the geometric inconsistency must be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth engaging with. It builds a targetless V2X roadside camera calibration method from known components—RAFT-style 4D correlation volumes, iterative flow refinement, and Mamba temporal fusion—and applies them to a real problem that existing single-vehicle calibration methods handle poorly. The ablations are honest: each component (temporal fusion, iteration, 4D volume) is shown to matter, and the comparison with LCCNet is fair. The code is released, and the results on two real datasets, including nighttime TUMTraf-V2X, are plausible. The reduction in standard deviation from 2.5° to 0.3° is a real robustness gain, not a cherry-picked mean error.\n\nThe biggest soft spot is the one the stress-test note identifies. Eqs. 1–3 define A as the vehicle LiDAR frame and treat TLC as a constant transform to the fixed roadside camera, which is geometrically inconsistent when the vehicle moves. The paper says data is \"aligned to a unified coordinate system\" and the vehicle is in motion, but never spells out that each point cloud is transformed by vehicle pose into the camera's world frame before projection. If the implementation does apply pose compensation, the equations are wrong and need correction. If it does not, the learned calibration flow would encode vehicle ego-motion confounded with camera rotation error, and the reported accuracy would be meaningless. This is the single most important thing the authors must clarify. I suspect it is a presentation issue—the method has been tested and works—but it is load-bearing enough that a referee should push hard on it.\n\nOther concerns are minor. The \"first\" claim is hedged and probably fine. Distance filtering before evaluation is a legitimate practical choice, but it means the headline numbers are conditional on being within 50 m; the authors acknowledge this and show distance curves, so it is not hidden. The synthetic-noise training distribution is standard for this literature, so I do not see it as circular.\n\nThis is a solid systems paper that deserves a serious referee. I would send it to peer review and ask for a revision that makes the coordinate frames explicit and repeats at least one experiment with the corrected equations. The central idea is useful and the evidence is strong enough to warrant the effort.","headline":"A genuinely useful V2X calibration pipeline with real-data results, but the frame bookkeeping for a moving vehicle's LiDAR is under-specified and must be clarified before the numbers can be trusted.","tokens_in":13031,"tokens_out":2188,"would_cite":true,"duration_ms":24137,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MamV2XCalib claims that a roadside camera's rotational misalignment can be recovered targetlessly from vehicle-side LiDAR point clouds by temporally fused calibration-flow regression, reaching about 0.63° mean rotation error on V2X-Seq.","keywords":["roadside camera calibration","targetless calibration","V2X","LiDAR-camera calibration","Mamba","state space model","4D correlation volume","infrastructure camera"],"falsifier":"If a LiDAR vehicle drives the same route past the same camera twice — once stationary and once at normal speed — and the network's predicted rotation differs by more than the reported error bars, or if calibration accuracy worsens monotonically with vehicle speed, then the missing ego-motion compensation is doing real work and the fixed-transform premise would fail. A simpler version: run inference on a sequence while the vehicle is parked with engine off; the method should still recover the same camera rotation if the premise holds.","tokens_in":11969,"feed_emoji":"📷","tokens_out":5310,"duration_ms":50521,"temperature":0.7,"pith_summary":"The paper tries to establish that fixed roadside cameras can be calibrated automatically while ordinary traffic flows, with no calibration targets, no reference images, and no road closures: a LiDAR-equipped vehicle drives past the camera, and the camera's rotational deviation from its initial mounting is recovered by a network. The motivation is that roadside cameras drift under weather and vibration, and uncalibrated rotation degrades vehicle-to-infrastructure fusion. The proposed system, MamV2XCalib, projects each vehicle point cloud into the camera using the flawed initial extrinsics, builds a 4D correlation volume between the resulting depth map and the RGB image, refines a pixel-level 'calibration flow' iteratively, and uses a Mamba state-space model to fuse these flow fields over time. On V2X-Seq with initial deviations up to ±20°, it reports mean rotation error 0.6313° with standard deviation 0.3211°, which the paper describes as an order-of-magnitude reduction in standard deviation over the single-vehicle baseline LCCNet.","feed_headline":"Passing cars can recalibrate roadside cameras without targets","feed_subtitle":"A V2X method fuses vehicle LiDAR with camera images to cut mean rotation error to 0.63 degrees.","key_machinery":"The central object is the 'calibration flow': a per-pixel correspondence field between the depth map formed by projecting the vehicle LiDAR point cloud with the inaccurate extrinsics and the roadside camera image. This flow is initialized at zero and refined by a RAFT-style iterative update: multi-scale 4D correlation volumes store all-pairs similarity between image and depth features, a GRU recurrently updates the flow using looked-up correlations and context features, and the sequence of flow fields from multiple frames and iterations is patched, position/temporally embedded, and passed through a bidirectional Mamba state-space model. A summarization token in the Mamba sequence is then regressed to the quaternion rotation error. The machinery's dual role is to make pixel-level matching robust under sparse, viewpoint-shifted point clouds and to let temporal fusion compensate for defective single frames.","core_discovery":"On the paper's own terms, the central discovery is that a cooperative V2X setup can invert the usual direction of assistance: the vehicle's LiDAR perception of the whole environment acts as a free calibration reference for the infrastructure camera. Because the camera is hinge-mounted, the only unknown that matters is the rotational deviation Rerror from the initial extrinsic Tinit, so the projection problem reduces to regressing a quaternion from the discrepancy between a projected depth map and the observed image. The paper argues that the key to making this work in V2X scenarios is temporal aggregation: individual frames are too sparse and viewpoint-shifted, but the camera's error is constant while the vehicle moves, so multiple flow estimates over time can be selectively fused by Mamba to suppress occlusions and bad frames. The resulting claim is a calibration method that needs no targets, no manual intervention, and no pre-built scene representation.","pith_inferences":["Because the paper never explicitly compensates for the vehicle's ego-motion when treating the LiDAR-to-camera transform as fixed, the learned network may implicitly absorb a vehicle-motion prior; a useful stress test would be calibrating from a stationary LiDAR vehicle or comparing estimates across different vehicle speeds.","The fixed-deviation-over-time assumption means temporal fusion can average away bad frames, but it also makes the method sensitive to time synchronization; misaligned timestamps could blur the fusion and bias the rotation estimate.","A testable extension is corridor-wide calibration: if a single LiDAR vehicle drives through a road network, every roadside camera that sees it could in principle be recalibrated in one pass, provided each camera's data stream is associated with the same vehicle trajectory."],"forward_implications":["Roadside camera recalibration could become a routine byproduct of normal traffic, since any LiDAR-equipped vehicle driving near a camera provides the data.","Calibration campaigns would no longer require placing checkerboards, closing lanes, or preparing reference images for every camera.","Large initial misalignments up to about ±20° can be brought back into the range where existing roadside perception systems operate correctly.","The reported stability (standard deviation reduced by roughly an order of magnitude versus LCCNet) implies fewer catastrophic calibration failures in fleet deployment.","The same architecture can be adapted to single-vehicle LiDAR-camera calibration when translation noise is also modeled."],"supporting_citations":[{"why":"Supplies the recurrent all-pairs field transform and iterative refinement design for the calibration flow.","marker":"[29]"},{"why":"Baseline single-vehicle LiDAR-camera cost-volume method that the paper compares against and whose 3D feature matching it replaces.","marker":"[21]"},{"why":"Provides the selective state-space model used to fuse flow fields across time and iterations.","marker":"[8]"},{"why":"Main real-world cooperative dataset used for training and evaluation, including the ±20° calibration benchmark.","marker":"[37]"},{"why":"Second real-world dataset used to test generalization, including nighttime scenes.","marker":"[43]"},{"why":"Train-free calibration method shown to fail in the large-deflection V2X scenario, serving as a failure-mode comparison.","marker":"[20]"},{"why":"Spatiotemporal state-space architecture that the temporal fusion design draws on.","marker":"[15]"}],"fun_headline_variants":["Roadside cameras calibrate as LiDAR cars drive by","V2X targetless calibration uses passing vehicle LiDAR","Automatic V2X camera calibration cuts error to 0.63°","No targets: passing LiDAR cars recalibrate road cams","Car LiDAR replaces calibration targets for road cameras"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method treats the LiDAR-to-camera transform as a fixed matrix even though the LiDAR is moving with the vehicle, and the paper does not describe transforming each point cloud into a common frame using vehicle pose; if vehicle motion is not compensated or learned implicitly, the projected depth maps encode motion as well as camera error.","fun_headline_variants_meta":{"raw":{"variants":["Roadside cameras calibrate as LiDAR cars drive by","V2X targetless calibration uses passing vehicle LiDAR","Automatic V2X camera calibration cuts error to 0.63°","No targets: passing LiDAR cars recalibrate road cams","Car LiDAR replaces calibration targets for road cameras"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1460,"prompt_tokens":994,"completion_tokens":466,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":381}},"tokens_in":610,"tokens_out":466,"duration_ms":4997,"temperature":1.0,"reasoning_tokens":381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:33:15.373424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a LiDAR vehicle drives the same route past the same camera twice — once stationary and once at normal speed — and the network's predicted rotation differs by more than the reported error bars, or if calibration accuracy worsens monotonically with vehicle speed, then the missing ego-motion compensation is doing real work and the fixed-transform premise would fail. A simpler version: run inference on a sequence while the vehicle is parked with engine off; the method should still recover the same camera rotation if the premise holds.","supporting_citations":[{"cited_title":"Lccnet: Lidar and camera self-calibration using cost volume network","cited_arxiv_id":null,"evidence_quote":"Baseline single-vehicle LiDAR-camera cost-volume method that the paper compares against and whose 3D feature matching it replaces."},{"cited_title":"V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecast- ing","cited_arxiv_id":null,"evidence_quote":"Main real-world cooperative dataset used for training and evaluation, including the ±20° calibration benchmark."},{"cited_title":"Tum- traf v2x cooperative perception dataset","cited_arxiv_id":null,"evidence_quote":"Second real-world dataset used to test generalization, including nighttime scenes."},{"cited_title":"Calib-Anything: Zero-training LiDAR-Camera Extrinsic Calibration Method Using Segment Anything","cited_arxiv_id":"2306.02656","evidence_quote":"Train-free calibration method shown to fail in the large-deflection V2X scenario, serving as a failure-mode comparison."},{"cited_title":"Videomamba: State space model for efficient video understanding","cited_arxiv_id":null,"evidence_quote":"Spatiotemporal state-space architecture that the temporal fusion design draws on."}],"review_version":1}