{"id":"83cb069a-d682-41f8-93d4-ac4c20e71da2","arxiv_id":"2411.14927","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"LET-VIC is an end-to-end lidar framework for vehicle-infrastructure cooperative detection and tracking that fuses temporal and multi-view features and learns to compensate calibration errors, outperforming the tested baselines on V2X-Seq-SPD.","lead":"A new system called LET-VIC combines lidar data from a self-driving car and roadside sensors to detect and track objects over time, while also correcting for sensor misalignment. In tests on a real-world vehicle-infrastructure dataset, it reports large accuracy gains over several existing methods, though the code is not yet released.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline gains over V2VNet, FFNet, and PointPillars rest on self-reimplemented baselines whose training details are not reported; if these baselines are undertuned, the reported margins are inflated.","rationale":"The reader identified baseline fairness as the weakest assumption, and I agree. I checked other potential weaknesses: the CEC ablation at 2Hz/64-dim is a weaker demonstration, but the core claim of outperforming baselines does not depend on that ablation. The strongest claim is vulnerable at the baselines, because the source of a 13+ point gain over PointPillars+AB3DMOT is unclear without training details. Even a well-designed architecture can appear to outperform poorly tuned reimplementations. I do not see an internal inconsistency in the proposed method, but the evidence is insufficient to accept the quantitative margins as stated. Keeping the CONDITIONAL verdict with code/config release and baseline parity checking is appropriate.","tokens_in":18288,"tokens_out":6092,"duration_ms":61121,"concrete_test":"Obtain the authors' training configurations, checkpoints, and evaluation scripts for all four baselines, and rerun Table I on the same V2X-Seq-SPD-10Hz split. If official V2VNet/FFNet code is used, verify equivalence to the authors' fusion and hyperparameters; for PointPillars+AB3DMOT, perform a validation-set search over detection score threshold, gating distance, and min-hits before reporting final numbers. The concern is settled only if the re-run reproduces all compared metrics within a small tolerance, e.g., ±1 mAP/AMOTA, using the released settings; otherwise the headline margins must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A3 says the authors 'also develop and reproduce' LET-V, V2VNet, FFNet, and PointPillars+AB3DMOT, but the paper reports no optimizer, learning rate, batch size, epochs, data augmentation, point-cloud range, detection confidence thresholds, or AB3DMOT association parameters for any baseline. The main numerical claim in Table I is a comparison against these self-implemented versions: LET-VIC obtains +15.0 mAP/+17.3 AMOTA over LET-V and +13.7 mAP/+13.1 AMOTA over PointPillars+AB3DMOT. Because V2VNet, FFNet, and PointPillars are used as generic frameworks rather than as released checkpoints, their performance depends heavily on the exact fusion scheme, training schedule, and tracking post-processing; no evidence is given that each baseline was tuned to convergence or that AB3DMOT parameters were optimized. The 'All code will be open-sourced' statement appears without a link or anonymous artifact, so the margins cannot currently be reproduced. This makes the central outperformance claim conditional on a hidden assumption of baseline parity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"LET-VIC is an end-to-end LiDAR-based detection-and-tracking framework for vehicle-infrastructure cooperation. The method extracts separate BEV features from vehicle and infrastructure point clouds with PointPillars, fuses them through a deformable VIC cross-attention module, and introduces a Calibration Error Compensation (CEC) module that learns per-reference-point offsets to correct misalignment between the two coordinate frames. Detection and tracking are performed jointly with TrackFormer-style queries. The paper reports experiments on the V2X-Seq-SPD-10Hz and V2X-Sim datasets, showing consistent improvements over a single-view end-to-end variant (LET-V) and over self-implemented tracking-by-detection baselines (V2VNet, FFNet, PointPillars+AB3DMOT) at 0-300 ms simulated latency. An ablation on a reduced 2Hz/64-dim setting indicates that CEC improves mAP and AMOTA.","tokens_in":18510,"tokens_out":6713,"duration_ms":56748,"significance":"If the quantitative claims are taken at face value, LET-VIC would be a useful step for cooperative perception: it is a clean end-to-end formulation that combines temporal queries with cross-sensor BEV fusion, it is evaluated on a real-world V2X sequential dataset, and its latency-robustness results in Table I address a practically important issue. The CEC idea, learning spatial offsets within the attention module rather than relying on post-hoc calibration, is simple and potentially transferable. The paper is also transparent about model complexity (514.86 GFLOPs, 30.11 M parameters). However, because the central comparison depends on self-reimplemented baselines whose training settings are not reported, and because the CEC ablation is not performed on the full model, the evidence as currently presented is conditional. The strengths above are real but do not yet verify the headline margins.","major_comments":[{"comment":"The four baseline methods (LET-V, V2VNet, FFNet, PointPillars+AB3DMOT) are described as \"developed and reproduced\" by the authors, but the paper gives no training hyperparameters (optimizer, learning rate, batch size, epochs, augmentation), no detection confidence thresholds, and no AB3DMOT association parameters. The main numerical claims in Table I and the text (+15.0 mAP/+17.3 AMOTA over LET-V and +13.7 mAP/+13.1 AMOTA over PointPillars at 0 ms) therefore rest on the unstated assumption that the re-implementations are trained as well as the published methods they are named after. Please provide complete baseline training details, or use official released checkpoints with known settings, and state whether any tuning was performed on the validation split.","section":"§IV-A3, Table I"},{"comment":"The CEC ablation is run on V2X-Seq-SPD-2Hz with the feature dimension reduced from 256 to 64, not on the full 10Hz/256-dim configuration of Table I. The text in Section IV-C3 concludes that \"LET-VIC with calibration error compensation\" improves performance, but this is only demonstrated for the reduced model. Please add a full-resolution CEC ablation (with and without CEC), or provide a concrete argument and supporting measurements showing that the reduced setting is representative of the full model.","section":"§IV-C2, Table III"},{"comment":"In the 0 ms columns, the V2VNet and FFNet rows are identical (mAP=0.349, AMOTA=0.377, AMOTP=1.258). This is suspicious and, as printed, makes the comparison with FFNet uninterpretable. Please correct or verify the entries, and check whether the 100-300 ms rows are also affected.","section":"Table I"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the comparisons. Since the baselines are re-implementations and the reported margins are often in the 5-15 point range, the reader cannot tell whether the differences are statistically meaningful. Please report mean±std over multiple seeds (or another justified significance measure) for at least the main results.","section":"§IV-B, Table I"},{"comment":"HYDRO-3D and MOT-CUP are discussed in the related work as cooperative tracking methods, yet neither appears in the experiments. Because the paper's contribution is precisely end-to-end temporal perception for VIC, a direct comparison with these methods should be included, or a concrete reason for their exclusion should be given. Otherwise the claim of outperforming \"representative\" cooperative temporal-perception baselines is not fully supported.","section":"§II-B, Table I"}],"minor_comments":[{"comment":"The vehicle-side grid is described as 512×512 with 0.2 m voxels over a 102.4 m range, but IV-A2 says the grid size is 0.512 m. These are inconsistent; please reconcile.","section":"§III-B vs §IV-A2"},{"comment":"The symbols 'xv mineh', 'xv maxeh', 'xi minnf', and 'xi maxnf' appear to be corrupted or missing subscripts; use x_veh^{min}, x_veh^{max}, etc. This makes the normalization in (6)-(7) hard to read.","section":"Equations (4)-(7)"},{"comment":"The contribution list calls the CEC approach \"self-supervised\", but the offsets in Section III-E are trained with the same detection/tracking supervision as the rest of the network; a term like \"learned\" or \"end-to-end trained\" would be more accurate.","section":"Section I, contribution list"},{"comment":"\"floating point operations per second (FLOPs)\" should be \"floating point operations\"; FLOPs is not a rate.","section":"Abstract and §IV-A2"},{"comment":"The table uses checkmark columns but never defines the baseline row (without CEC) besides the table title; also the reduced setting should be stated directly in the table caption to avoid confusion.","section":"Table III caption"},{"comment":"The availability statement \"All code will be open-sourced\" has no repository link or artifact identifier; if code is released, provide a URL or mention that it will be linked in the camera-ready version.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript has already been published (DOI 10.1109/JIOT.2025.3552526), so this review concerns the claims as written rather than novelty at submission. The most serious issues are fixable: full baseline specifications, a full-resolution CEC ablation, and correcting the duplicated V2VNet/FFNet numbers. I would not recommend rejection on the basis of the current evidence, but a major revision is needed to make the central claims verifiable. There is also a fit question: the paper is in an IoT journal but the contribution is mostly a computer-vision method; the communication-delay analysis partially addresses the IoT angle."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know before you read this one. First, it does something genuinely new: it builds an end-to-end LiDAR tracking model specifically for vehicle-infrastructure cooperation, using deformable cross-attention between vehicle and infrastructure BEV features, plus a calibration-error-compensation module that learns per-reference-point offsets in that attention. That combination is new, and the CEC idea is a real addition. Second, the main comparison table is less solid than it looks: the baselines (V2VNet, FFNet, PointPillars+AB3DMOT) were re-implemented by the authors with no reported training details, so the +13-17 point margins over them are conditional on those re-implementations being fair.\n\nWhat the paper does well: the task definition is clear, the method equations are explicit enough to reconstruct, and the experiments cover four latency settings plus a cross-dataset check on V2X-Sim. The CEC ablation, though run at reduced resolution, shows consistent gains across all latencies. The authors also report FLOPs and parameter count, which is helpful.\n\nSoft spots, in rough order. The baseline re-implementations are the biggest. Section IV-A3 says 'we also develop and reproduce four methods' but gives no optimizer, learning rate, epochs, augmentation, detection thresholds, or AB3DMOT parameters. The main numerical claims compare against these self-implemented versions. If the baselines are undertuned, the gains are inflated. The stress-test note is accurate on this. Second, the CEC ablation is on a 2Hz/64-dim model, not the full 10Hz/256-dim one, so we don't know if the effect holds at full scale. That is a moderate gap. Third, MOT-CUP and HYDRO-3D are cited but never compared, which weakens the 'state-of-the-art' claim. No error bars or significance tests, but the margins are big enough that I would not weight that heavily. The citation pattern leans on the authors' own prior work, but that is not a flaw by itself—the dataset is real and the self-citations are relevant.\n\nWho this is for: anyone in V2X cooperative perception who cares about tracking rather than single-frame detection. Worth reading for the CEC module and the end-to-end framing, but I would not trust the absolute numbers until the baselines are reproducible.\n\nRecommendation: send it to peer review—the core idea is sound and the contribution is real. The reviewers should ask for released code or at least full baseline training settings, a full-model CEC ablation, and a comparison against MOT-CUP/HYDRO-3D. I would make those conditions for acceptance.","headline":"A credible end-to-end VIC tracking framework with a useful calibration-compensation idea, but the headline gains rest on unverifiable baseline re-implementations, so the numbers should be read as provisional.","tokens_in":19086,"tokens_out":4912,"would_cite":true,"duration_ms":41899,"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":"LET-VIC introduces a LiDAR-based end-to-end tracking framework that fuses vehicle and roadside BEV features and uses learned calibration offsets, and it reports large, delay-robust gains over single-vehicle and tracking-by-detection…","keywords":["vehicle-infrastructure cooperation","end-to-end multi-object tracking","LiDAR perception","BEV feature fusion","cross-attention","calibration error compensation","temporal perception","V2X-Seq-SPD"],"falsifier":"Retrain V2VNet, FFNet, and PointPillars+AB3DMOT on the same V2X-Seq-SPD-10Hz split with the same training schedule, hyperparameters, augmentation, and compute budget used for LET-VIC; if LET-VIC no longer leads by roughly +13.7 mAP and +13.1 AMOTA at 0 ms, the central comparative claim fails.","tokens_in":18072,"feed_emoji":"🚗","tokens_out":8335,"duration_ms":70398,"temperature":0.7,"pith_summary":"LET-VIC is a proposed answer to a concrete failure: a self-driving car's LiDAR sees only part of the scene, and cooperative perception systems that fuse roadside data usually detect frame-by-frame and then track separately, leaving temporal context underused. The paper argues that a single end-to-end network, which fuses vehicle and infrastructure Bird's Eye View (BEV) features with cross-attention and carries track identity across frames, detects and tracks better than either single-vehicle end-to-end perception or cooperative tracking-by-detection. On the real-world V2X-Seq-SPD-10Hz validation set, LET-VIC reports mAP 0.606 and AMOTA 0.640 at zero latency, +15.0 mAP and +17.3 AMOTA over its single-vehicle counterpart, and at least +13.7 mAP and +13.1 AMOTA over V2VNet, FFNet, and PointPillars with AB3DMOT, with the advantage persisting under 100-300 ms communication delays. If true, this makes the case that temporal perception in cooperative driving should be treated as one learnable detection-and-tracking problem rather than detection followed by association, and that calibration error can be absorbed by learned reference-point offsets instead of precise extrinsic calibration alone.","feed_headline":"Roadside-plus-car LiDAR lifts tracking by 17 points","feed_subtitle":"An end-to-end V2X model beats single-vehicle and tracking-by-detection baselines on real data even at 300 ms delay.","key_machinery":"The load-bearing mechanism is the VIC Cross-Attention module, a deformable-attention fusion layer that takes the vehicle-side BEV queries, projects them onto both the vehicle and infrastructure feature maps, and then refines each projected reference point with a learned calibration offset ($\\Delta x$, $\\Delta y$) before sampling features around the corrected points. An infrastructure mask excludes projections that fall outside the roadside feature bounds. Working with the TrackFormer-style decoder of detect queries and track queries, this module is what lets LET-VIC fuse two LiDAR views and maintain object identity across frames in a single training pass.","core_discovery":"The authors introduce LET-VIC, which they describe as the first LiDAR-based end-to-end tracking framework for vehicle-infrastructure cooperation. The central claim is that detection and tracking should be solved jointly in one Transformer pipeline that fuses vehicle and infrastructure Bird's Eye View (BEV) features, rather than by detecting each frame separately and associating tracks afterwards. The VIC Cross-Attention module performs the fusion with deformable attention, and the Calibration Error Compensation (CEC) module learns per-reference-point offsets $\\Delta x$ and $\\Delta y$ to correct coordinate-transformation misalignment. On the real-world V2X-Seq-SPD-10Hz validation set, LET-VIC reports 0.606 mAP and 0.640 AMOTA at 0 ms latency, improving on its single-vehicle counterpart LET-V by +15.0 mAP and +17.3 AMOTA, and on the tracking-by-detection baselines V2VNet, FFNet, and PointPillars with AB3DMOT by at least +13.7 mAP and +13.1 AMOTA. The advantage persists under 100, 200, and 300 ms simulated communication delays, and the CEC ablation on the 2Hz dataset shows consistent mAP and AMOTA gains at every tested delay.","pith_inferences":["Editorial inference: if the margins survive controlled baseline tuning, the main design lesson for cooperative perception is that temporal end-to-end heads may matter more than better data association for tracking in V2X settings.","Editorial inference: the CEC module's learned offsets could be read as a soft, online estimate of extrinsic calibration error; one testable extension is to compare the learned offsets against measured misalignment values and use them as a calibration-drift monitor.","Editorial inference: because the cross-dataset gains on V2X-Sim are much smaller (+2.2 mAP, +6.7 AMOTA) under intermittent cooperation, the framework's robustness to dropped or absent infrastructure features is worth isolating explicitly with feature-dropout tests.","Editorial inference: the CEC gains are demonstrated only on the reduced 2Hz, 64-dimension configuration; verifying them at the full 10Hz, 256-dimension setting would determine whether the module's benefit scales with resolution."],"forward_implications":["If the reported results hold, cooperative temporal perception is better formulated as one end-to-end detection-and-tracking problem; the +15.0 mAP and +17.3 AMOTA gains over LET-V show that roadside LiDAR data materially extends what the ego vehicle can perceive.","End-to-end tracking alone is competitive even without cooperation: LET-V, using only ego-vehicle point clouds, approaches PointPillars+AB3DMOT and beats V2VNet and FFNet, suggesting that track queries carry useful temporal context that box-association post-processing does not.","Learned calibration compensation is a viable substitute for precise extrinsic alignment: adding CEC improves mAP and AMOTA at all latencies tested on the 2Hz setting, including +6.4 mAP at 0 ms.","The framework tolerates communication delay gracefully; at 300 ms delay it still outperforms the best baseline by +8.5 mAP and +9.2 AMOTA, so the fusion strategy is not brittle to stale infrastructure features."],"supporting_citations":[{"why":"It supplies the real-world V2X-Seq-SPD dataset and its 10Hz and 2Hz re-extractions used for all main comparisons and ablations.","marker":"[9]"},{"why":"It provides the PointPillars LiDAR feature backbone for both vehicle and infrastructure branches and is also one of the tracking-by-detection baselines.","marker":"[53]"},{"why":"It supplies the end-to-end Transformer detection-and-tracking formulation with detect and track queries that LET-VIC adapts.","marker":"[6]"},{"why":"It provides the TrackFormer joint detection-and-tracking decoder design that informs the tracking head.","marker":"[5]"},{"why":"It is the cooperative detection baseline paired with AB3DMOT that LET-VIC must outperform.","marker":"[18]"},{"why":"It is the vehicle-infrastructure cooperative detection baseline paired with AB3DMOT in the comparison.","marker":"[22]"},{"why":"It provides the AB3DMOT tracker used with the tracking-by-detection baselines, defining that comparison paradigm.","marker":"[58]"},{"why":"It supplies the simulated V2X-Sim dataset used to test generalization to different cooperation rates and sensor configurations.","marker":"[32]"}],"fun_headline_variants":["End-to-end V2X LiDAR tracking: +17% AMOTA over single car","Cooperative LiDAR tracker beats single-vehicle by 17 points","Infrastructure+car LiDAR fusion lifts tracking up to 17.3%","End-to-end V2X tracking outranks baselines, robust to delays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison against V2VNet, FFNet, and PointPillars assumes those re-implemented baselines were trained as carefully and with as much compute as LET-VIC; the paper gives no training details for them, so if they were undertuned the reported margins are inflated.","fun_headline_variants_meta":{"raw":{"variants":["End-to-end V2X LiDAR tracking: +17% AMOTA over single car","Cooperative LiDAR tracker beats single-vehicle by 17 points","Infrastructure+car LiDAR fusion lifts tracking up to 17.3%","End-to-end V2X tracking outranks baselines, robust to delays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000939,"raw_usage":{"total_tokens":4095,"prompt_tokens":1106,"completion_tokens":2989,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":722,"completion_tokens_details":{"reasoning_tokens":2903}},"tokens_in":722,"tokens_out":2989,"duration_ms":21920,"temperature":1.0,"reasoning_tokens":2903,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:41:54.464870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain V2VNet, FFNet, and PointPillars+AB3DMOT on the same V2X-Seq-SPD-10Hz split with the same training schedule, hyperparameters, augmentation, and compute budget used for LET-VIC; if LET-VIC no longer leads by roughly +13.7 mAP and +13.1 AMOTA at 0 ms, the central comparative claim fails.","supporting_citations":[{"cited_title":"V2vnet: Vehicle-to-vehicle communication for joint percep- tion and prediction,","cited_arxiv_id":null,"evidence_quote":"It is the cooperative detection baseline paired with AB3DMOT that LET-VIC must outperform."},{"cited_title":"Flow-based feature fusion for vehicle-infrastructure cooperative 3d object detection,","cited_arxiv_id":null,"evidence_quote":"It is the vehicle-infrastructure cooperative detection baseline paired with AB3DMOT in the comparison."},{"cited_title":"3d multi-object tracking: A baseline and new evaluation metrics,","cited_arxiv_id":null,"evidence_quote":"It provides the AB3DMOT tracker used with the tracking-by-detection baselines, defining that comparison paradigm."}],"review_version":1}