{"id":"8d59d67b-1c71-4117-8ca9-02b4bd7477a2","arxiv_id":"2502.01357","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Bayes-4DRTrack combines transformer-based motion prediction, Monte Carlo dropout uncertainty, and Doppler-based data association to improve 3D multi-object tracking with 4D radar on the K-Radar dataset.","lead":"This paper presents a radar-based system for tracking multiple vehicles in 3D that predicts where vehicles will move next using a transformer network, and estimates how uncertain those predictions are by running the network multiple times with random dropouts. It also uses vehicle speed from radar to match detections across frames, and reports better tracking accuracy than several standard Kalman-filter baselines on the K-Radar dataset, which includes fog, rain, and snow.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MC Dropout variance in Eq. (2) is used as process noise without any calibration check; the marginal gain from this component is only 0.003 AMOTA, so the paper overstates the role of adaptive noise modeling in the reported 5.7% gain.","rationale":"The reader identified the MC Dropout variance calibration as the weakest assumption; I agree that this is a real gap, but I would nuance the diagnosis. The headline 5.7 percentage-point gain (0.485 -> 0.542) is not predominantly driven by the adaptive process-noise mechanism that the paper emphasizes. Table II shows that adding the prediction-side variance (V) to the already-Bayesian detection system changes AMOTA by only +0.003 (rows 9->10), whereas detection-side Loss Attenuation and MC Dropout together account for +0.045 of the gain relative to the non-Bayesian baseline (rows 1->7). Thus a calibration failure in Eq. (2) would not collapse the reported AMOTA; it would, however, undermine the paper's central claim that Bayesian approximation in prediction yields meaningful gains through adaptive noise modeling, which is the paper's stated novelty. The test I propose settles this by comparing against a tuned fixed Q. I also note Table II's 'Baseline (CV)' row has identical metrics to Table I's Prediction Network n=3 Two-stage row; if 'CV' means constant-velocity, this is an inconsistency, but if it means fixed covariance it is merely an ambiguous label. The reader's conditional verdict remains appropriate; no fatal flaw is present in the reported numbers based on the text alone.","tokens_in":8921,"tokens_out":14625,"duration_ms":125820,"concrete_test":"Re-run the full system (transformer prediction + two-stage association + detection MC Dropout/Loss Attenuation) but replace the MC Dropout variance in Eq. (2) with a fixed, diagonal process noise covariance Q whose magnitude is tuned on the K-Radar validation set (e.g., grid search over a scalar multiplier of the identity or of the position diagonal). If the best fixed-Q configuration achieves AMOTA >= 0.542 on the test set, the adaptive Bayesian variance is not responsible for the reported gain and the central prediction-stage contribution claim fails. If fixed-Q plateaus below 0.542, the variance conveys useful information, though calibration still needs separate checking.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is that Bayesian approximation in the prediction stage—specifically, using the MC Dropout sample variance from Eq. (2) as the Kalman filter's process noise covariance—improves tracking. Yet this component contributes only 0.003 AMOTA (0.539 to 0.542 in Table II, rows 9–10); the 5.7 percentage-point headline gain over the CV baseline (0.485) is overwhelmingly due to detection-side Loss Attenuation plus detection MC Dropout (rows 1 to 7: +0.045) and to the transformer prediction network itself (+0.007). The ablation does not isolate or validate the adaptive-noise mechanism. No training details for the prediction network are given (loss function, dropout rate, whether it was trained with a Gaussian likelihood), and no calibration analysis (e.g., reliability diagram) is reported to show Eq. (2) is a meaningful covariance. Without this, the paper's claim that 'properly modeling uncertainty in the prediction step can meaningfully boost tracking performance' (Section IV-B.3) is unsupported; the small gain from V could simply reflect a chance improvement from a particular schedule of Q values, not Bayesian uncertainty. This is the load-bearing vulnerability of the novelty claim, even though the empirical AMOTA values themselves are not shown to be wrong.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Bayes-4DRTrack, a 3D multi-object tracking system for 4D radar built on a tracking-by-detection pipeline. The detector is RTNH with MC Dropout and Loss Attenuation; the prediction stage is a transformer-based motion prediction network with MC Dropout, whose sample variance is used as an adaptive process-noise covariance for the Kalman filter; and data association combines Mahalanobis distance with a Doppler-based velocity affinity. The method is evaluated on the K-Radar dataset, reporting an AMOTA of 0.542 against 0.485 for a constant-velocity Mahalanobis baseline, described as a 5.7% gain. Ablations examine input horizon, two-stage association, and the contributions of Bayesian approximation on the detection and prediction sides.","tokens_in":9169,"tokens_out":6464,"duration_ms":61497,"significance":"If the reported results are statistically reliable, Bayes-4DRTrack would be a useful demonstration that 4D radar Doppler information and learned motion models can improve tracking in adverse weather. The paper has notable strengths: the use of a common detector (RTNH) across all compared tracking methods is a fair-control design; the formulas for Loss Attenuation, MC Dropout averaging, and the association cost are standard and correctly stated; and the ablation study is structured to separate detection-side and prediction-side contributions. However, the headline claim is ambiguous (percentage points vs. relative percent), and the specific contribution of the adaptive prediction variance—which the paper presents as a core novelty—is very small in the ablation (about 0.003 AMOTA) and is reported without error bars or significance tests. The paper also omits key experimental details needed to assess reproducibility, including the dataset split, training protocol for the prediction network, and several hyperparameter values. The contribution is therefore plausible but not yet substantiated at the level claimed.","major_comments":[{"comment":"The central claim that using MC Dropout variance from the prediction step as process noise 'meaningfully boosts tracking performance' is not supported by the ablation. In Table II, rows 7–10, the baseline with detection-side MC Dropout and Loss Attenuation achieves 0.537; adding MC Dropout prediction gives 0.539; adding the estimated variance V gives 0.542. The marginal contribution of V is therefore 0.003 AMOTA (and 0.001 in rows 2–3 and 0.002 in rows 5–6). These differences are within the range that could arise from the stochasticity of MC Dropout alone, and the paper reports no repeated runs, error bars, or significance tests. The authors should either provide statistical evidence that the V component is a real improvement or reframe the paper's contribution away from adaptive prediction noise.","section":"Section IV-B.3, Table II"},{"comment":"The abstract's '5.7% gain' is an absolute percentage-point difference: 0.542 minus 0.485 equals 0.057, not a 5.7% relative improvement, which would be about 11.8%. The same issue appears in Section IV.C, where a '4.5%' gain over SimpleTrack is also absolute percentage points (0.542 vs. 0.497). The text should consistently say 'percentage points' and should clearly state the selected baseline; against the stronger SimpleTrack baseline in Table III the gain is 4.5 percentage points, not 5.7. This matters because the comparison baseline affects how the contribution is perceived.","section":"Abstract and Section IV.C"},{"comment":"The manuscript does not provide enough experimental detail to reproduce or fully assess the prediction network. There is no description of the transformer architecture, the loss function, the dropout rate or placement, the optimizer, the number of training epochs, or the train/validation/test split of K-Radar. In particular, if input horizon n=3 and other hyperparameters were selected on the same data used for the final evaluation, the comparisons in Table I may overstate the advantage. The authors should report the evaluation protocol and training split, and ideally provide code or a detailed configuration.","section":"Section III-C and Section IV-B"},{"comment":"The MC Dropout predictive variance from Eq. (2) is used as the Kalman filter's process noise covariance with no calibration check. A raw sample variance over NP stochastic forward passes is only meaningful as a covariance if the network's uncertainty is well calibrated; the paper provides no reliability diagrams, coverage analysis, or comparison of predicted variance against actual trajectory errors. Since the entire adaptive-noise component rests on this variance, the authors should add a calibration study or at least verify that the estimated variance is correlated with prediction error, rather than simply reporting a small AMOTA change.","section":"Section III-C, Eq. (2) and Section IV-B.3"}],"minor_comments":[{"comment":"The meaning of 'V' in the table header is ambiguous: in rows 3, 6, 10, and 12 it appears to denote prediction variance used as process noise, while in rows 8 and 11 it denotes detection variance used as measurement noise. The table should use distinct labels, such as V_pred and V_det, to make the ablation interpretable at a glance.","section":"Table II"},{"comment":"The two-stage association description is internally inconsistent with the final cost formula. The text says Stage 1 uses Mahalanobis distance and Stage 2 refines unmatched tracks with Doppler, but then defines a final association cost A(i,j) = w1*DM(i,j) + w2*(1-AR(i,j)), which suggests a combined cost for all pairs. Please clarify whether the matching is sequential or joint, and specify the matching algorithm (e.g., Hungarian or greedy) and any gating thresholds.","section":"Section III-D"},{"comment":"The Doppler affinity term AR(i,j) uses a velocity difference Δv_ij but does not define whether this is a full 3D velocity vector, a radial Doppler velocity, or a projected component. Since 4D radar provides radial velocity, the definition should be explicit, and the values of w1, w2, and the Doppler sigma should be reported.","section":"Section III-D"},{"comment":"Reference [32] is cited for the AMOTA/AMOTP metrics, but the cited paper (Henriques et al., ECCV 2012) is about circulant tracking kernels. The appropriate citation is likely Weng et al. [10] or the nuScenes benchmark where AMOTA is defined.","section":"References"},{"comment":"Eq. (2) uses the sample variance with a 1/NP normalization, but it is not stated whether the predicted states p_i include the orientation θ and how the variance for θ is combined with the positional variances when forming the Kalman process-noise matrix. Please specify the state vector and the exact mapping from (μ, σ^2) to Q.","section":"Section III-C"},{"comment":"Table I explores input horizons n=2,...,5, but the paper does not explain how tracks shorter than n frames are initialized or handled during the first frames (e.g., whether a constant-velocity fallback is used). This detail is needed to interpret the performance at different horizons.","section":"Section IV-B.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's main novelty is presented as Bayesian approximation in the prediction stage with adaptive process noise, but the ablation evidence for that component is thin. The empirical numbers could still be useful to the radar-MOT community if the authors add statistical validation and calibrate the uncertainty claim. I would not recommend rejection if the authors honestly reframe the contribution and supply the missing experimental details. The use of the authors' own K-Radar dataset and RTNH detector is acceptable given the common-detector baseline setup, but the lack of a stated split and training protocol is a concern. The paper is a borderline contribution that needs a solid revision to be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the punchline: this is a plausible engineering integration of known pieces — MC Dropout, loss attenuation, a transformer motion predictor, and Doppler-gated association — for 4D radar MOT on K-Radar. What's actually new is the specific combination, not any single component. The paper is clearly written and the ablations in Table II are more informative than most, which makes the soft spots easier to see.\n\nWhat it does well: the formulas for MC Dropout averaging and loss attenuation are standard and correctly stated. Using the same detector (RTNH) across all comparison methods is the right call for a tracking-focused study. The ablation table breaks down detection-side and prediction-side contributions, which is exactly what a reviewer wants.\n\nNow the real problem. The abstract and Section IV-B.3 attribute the AMOTA gain to 'properly modeling uncertainty in the prediction step,' but the prediction-side variance (V) is worth only 0.003 AMOTA (0.539 to 0.542 in Table II). The 5.7 percentage-point gain over the CV baseline is almost entirely from the detection-side MC Dropout + Loss Attenuation (+0.045) plus the transformer itself (+0.007). So the marginal contribution of the adaptive process-noise mechanism, which the paper sells as a core innovation, is neither isolated nor validated. There is no calibration check (e.g., reliability diagram) showing that Eq. (2) is a meaningful covariance; the tiny gain could easily come from tweaking the Q schedule. The claim that Doppler association reduces ID switches is also not supported: in Table I, IDS actually goes up slightly for the CV model with two-stage association (259 to 261).\n\nMinor issues: '5.7% gain' is ambiguous (it is 5.7 percentage points), and all numbers are single-run with no error bars or significance tests. The authors select the input horizon and the 'estimated variance only in prediction' choice on the same K-Radar test data, so there is a selection-snooping risk. No code or training details for the prediction network are given.\n\nWould this get a serious referee? Yes. The system-level result (0.542 AMOTA) is competitive and the paper is mostly transparent about its ablations. But the load-bearing novelty claim needs to be tested properly: run multiple seeds, report variance, and ideally provide a calibration analysis of the predictive covariance. Without that, the paper should be revised to claim only what the data support: detection-side Bayesian approximation and the transformer motion model carry the gains.\n\nWho is this for? Researchers in radar-based tracking or practitioners building robust AD perception stacks. It's a useful reference point, but not a foundational result. If I were an area chair, I'd send it to reviewers with a request for a major revision, focusing on the variance ablation and error bars.","headline":"Useful integration of known uncertainty tricks for 4D radar MOT, but the headline gain mostly comes from detection-side changes, not the advertised adaptive process noise.","tokens_in":9734,"tokens_out":2892,"would_cite":true,"duration_ms":24696,"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":"Bayes-4DRTrack claims that replacing fixed Kalman covariance with MC Dropout variance and adding Doppler association lifts AMOTA on K-Radar from 0.485 to 0.542.","keywords":["4D radar","multi-object tracking","Bayesian approximation","MC Dropout","loss attenuation","transformer motion prediction","Doppler data association","K-Radar"],"falsifier":"Measure the calibration of the MC Dropout variance on K-Radar: compute the empirical squared error of $\\mu_{\\hat{p}_{t+1}}$ against the true next state and compare it with $\\sigma^2_{\\hat{p}_{t+1}}$ across tracks, for instance by binning predictions by predicted variance and checking that observed error matches. If the variance is systematically miscalibrated, or if permuting variance values across tracks (removing per-track information) leaves AMOTA at 0.542, then the reported gain is not caused by per-track uncertainty estimation.","tokens_in":8682,"feed_emoji":"📡","tokens_out":7225,"duration_ms":54527,"temperature":0.7,"pith_summary":"Bayes-4DRTrack claims that a 4D-radar multi-object tracker can be made markedly more accurate by treating uncertainty as a learned quantity rather than a fixed Kalman covariance. The paper's central empirical claim is that combining MC Dropout and Loss Attenuation in both detection and trajectory prediction, together with a transformer motion model and Doppler-based two-stage data association, raises AMOTA on the K-Radar dataset to 0.542, against 0.485 for a constant-velocity Mahalanobis baseline. The largest single source of the gain is using the variance of MC Dropout trajectory samples as the Kalman filter's process noise, an adaptive replacement for fixed noise covariance. A sympathetic reader would care because 4D radar is one of the few sensors that stays reliable in fog, rain, and snow, so improved radar tracking accuracy directly addresses the adverse-weather case where camera and LiDAR systems degrade.","feed_headline":"Bayesian uncertainty lifts 4D radar tracking to 0.542 AMOTA","feed_subtitle":"MC Dropout variance replaces fixed Kalman noise while Doppler association separates close targets.","key_machinery":"The central mechanism is a Kalman filter whose noise is not fixed. In the prediction step, $N_P=10$ MC Dropout forward passes on a transformer motion model produce predicted states $\\hat{p}^i_{t+1}$; their empirical mean $\\mu_{\\hat{p}_{t+1}}$ becomes the predicted state and their empirical variance $\\sigma^2_{\\hat{p}_{t+1}}$ (Eq. 2) becomes the process noise covariance. Detection uses the same idea: $N_D=10$ MC Dropout samples, IoU clustering, and Loss Attenuation (Eq. 1) that weights the regression loss by an input-dependent observation noise $\\sigma(x_i)$. Matching runs in two stages, first Mahalanobis distance with the dynamically updated covariance, then a Doppler velocity affinity $A_R(i,j)=\\exp(-\\|\\Delta v_{ij}\\|^2 / 2\\sigma^2)$, so objects with similar positions but different radial velocities are separated.","core_discovery":"The paper's core discovery, stated on its own terms, is that a tracking-by-detection 3D MOT system for 4D radar reaches 0.542 AMOTA on K-Radar by applying Bayesian approximation, via MC Dropout and Loss Attenuation, to both the RTNH detector and a transformer-based motion predictor; using the sample variance of $N_P=10$ MC Dropout trajectory predictions as the Kalman filter's process noise covariance; and adding a Doppler-velocity association stage after Mahalanobis matching. The ablation results attribute the largest gain to this adaptive prediction variance: with Loss Attenuation in detection, MC Dropout prediction plus estimated variance moves AMOTA from 0.515 to 0.542, whereas feeding detection variance into measurement noise slightly lowers it to 0.537. The full system also reports the best true-positive count (11,682) and the fewest identity switches (246) among the methods compared in Table III.","pith_inferences":["If the MC Dropout variance is calibrated, the same adaptive-noise recipe could transfer to camera- or LiDAR-based trackers, whose Kalman filters also use fixed covariances; the paper does not test that transfer.","The Doppler association stage is likely to matter most in dense traffic with many similar-position targets at different speeds; a stress test on highway scenes with high relative velocities would isolate its contribution.","A direct calibration check of Eq. (2), comparing predicted variance against empirical squared error on held-out K-Radar tracks, would tell whether the AMOTA gain comes from well-calibrated uncertainty or merely from a variance schedule that happens to help the filter.","The asymmetry reported in the ablation, where prediction variance helps and detection variance hurts, suggests that over-trusting noisy detections is worse than being conservative in motion; the paper asserts this asymmetry but does not explain it."],"forward_implications":["Replacing fixed process noise with MC-Dropout-derived variance raises AMOTA from 0.495 to 0.542 in the full configuration, indicating the adaptive variance is the largest single contributor.","Doppler-based two-stage association consistently increases true positives and decreases false positives relative to Mahalanobis-only matching in the reported ablation.","Loss Attenuation in detection alone lifts the constant-velocity baseline from 0.492 to 0.515 AMOTA.","With a three-frame input horizon the transformer motion model outperforms the constant-velocity model; longer horizons of four or five frames degrade AMOTA.","The full Bayes-4DRTrack outperforms AB3DMOT, Probabilistic 3D MOT, and SimpleTrack on K-Radar, beating the strongest of those baselines by 4.5 AMOTA points."],"supporting_citations":[{"why":"Supplies the K-Radar dataset and the RTNH 4D radar detector used as the base detection backbone.","marker":"[2]"},{"why":"Provides the transformer architecture on which the motion prediction network is built.","marker":"[9]"},{"why":"Defines the AB3DMOT baseline and the 3D MOT metrics used for comparison.","marker":"[10]"},{"why":"Serves as the probabilistic 3D MOT baseline that also addresses uncertainty, against which Bayes-4DRTrack is compared.","marker":"[11]"},{"why":"Provides the SimpleTrack baseline and the life-cycle management procedure adopted for track initialization and removal.","marker":"[12]"},{"why":"Establishes MC Dropout as a Bayesian approximation method, which the paper uses to generate predictive variance in detection and prediction.","marker":"[14]"},{"why":"Provides the Loss Attenuation approach for heteroscedastic uncertainty weighting used in the detection loss.","marker":"[15]"}],"fun_headline_variants":["Bayesian noise model boosts 4D radar tracking to 0.542 AMOTA","MC Dropout variance lifts 4D radar tracking to 0.542 AMOTA","Adaptive Kalman noise from Bayes nets improves radar MOT to 0.542","Doppler association and Bayesian variance push radar AMOTA to 0.542","4D radar tracker uses Bayesian uncertainty to reach 0.542 AMOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the sample variance of the $N_P=10$ MC Dropout trajectory predictions is a calibrated measure of how wrong the predicted position is, so feeding it into the Kalman filter as process noise genuinely adapts the filter rather than merely rescaling its gains.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian noise model boosts 4D radar tracking to 0.542 AMOTA","MC Dropout variance lifts 4D radar tracking to 0.542 AMOTA","Adaptive Kalman noise from Bayes nets improves radar MOT to 0.542","Doppler association and Bayesian variance push radar AMOTA to 0.542","4D radar tracker uses Bayesian uncertainty to reach 0.542 AMOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1687,"prompt_tokens":901,"completion_tokens":786,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":680}},"tokens_in":517,"tokens_out":786,"duration_ms":6096,"temperature":1.0,"reasoning_tokens":680,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:33:36.744265+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the calibration of the MC Dropout variance on K-Radar: compute the empirical squared error of $\\mu_{\\hat{p}_{t+1}}$ against the true next state and compare it with $\\sigma^2_{\\hat{p}_{t+1}}$ across tracks, for instance by binning predictions by predicted variance and checking that observed error matches. If the variance is systematically miscalibrated, or if permuting variance values across tracks (removing per-track information) leaves AMOTA at 0.542, then the reported gain is not caused by per-track uncertainty estimation.","supporting_citations":[{"cited_title":"K-radar: 4d radar object detection for autonomous driving in various weather conditions,","cited_arxiv_id":null,"evidence_quote":"Supplies the K-Radar dataset and the RTNH 4D radar detector used as the base detection backbone."},{"cited_title":"Probabilistic 3d multi- modal, multi-object tracking for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Serves as the probabilistic 3D MOT baseline that also addresses uncertainty, against which Bayes-4DRTrack is compared."}],"review_version":1}