{"id":"6c296228-36a1-47ea-ab5e-6b879943b76e","arxiv_id":"2507.03441","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Radar Tracker adds temporal offset prediction and attention-based appearance association to a radar instance segmentation backbone, achieving an LSTQ of 66.8 on the RadarScenes moving-instance tracking benchmark.","lead":"This paper presents Radar Tracker, a learning-based method that tracks moving objects in sparse, noisy automotive radar point clouds by combining motion-based center association with appearance-based attention matching. It reports the best moving-instance tracking scores on the RadarScenes benchmark, which matters for autonomous vehicles that need reliable perception in rain, fog, and snow.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim rests on excluding Eq-4D-StOP without a quantitative comparison; if that method transfers to RadarScenes, Table I's 66.8 LSTQ may not be SOTA.","rationale":"The most load-bearing condition for the central claim is that no excluded method outperforms the reported 66.8 LSTQ. The paper explicitly excludes Eq-4D-StOP, described as the best-performing 4D panoptic method, based on an unquantified argument about rotation augmentation being detrimental to radar data. That exclusion directly threatens the 'state-of-the-art' claim. The reader's weakest_assumption focused on temporal-offset regression targets for single-point instances; that is a legitimate modeling concern, but it is secondary because the paper's empirical results on RadarScenes would still stand even if those targets are noisy. The reader's rationale, however, already identifies the missing strongest baseline and single-run evaluation, so the final conditional verdict is appropriate. The proposed concrete check would settle whether the exclusion is justified and whether the headline claim holds; if Eq-4D-StOP can reach or exceed the reported number, the verdict should move toward rejection of the SOTA claim specifically, otherwise the current conditional acceptance remains.","tokens_in":12054,"tokens_out":8757,"duration_ms":110034,"concrete_test":"Run Eq-4D-StOP [59] on the same RadarScenes split (130/6/22) in two configurations: (a) as released, with its original rotation augmentation, and (b) with rotation augmentation disabled or replaced by radar-compatible augmentation. Report LSTQ and S_assoc on the 22 test sequences. If either configuration reaches or exceeds 66.8 LSTQ / 48.2 S_assoc, the SOTA claim is refuted. Also run the proposed Radar Tracker with 3 random seeds and report mean±std LSTQ to verify the 7.5-point advantage over center tracking is outside run-to-run noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract and Sec. I is explicitly comparative: 'state-of-the-art performance ... compared to the current state of the art.' The evidence in Table I supports that claim only among the selected baselines (MOT, center tracking, CA-Net). Sec. IV-B states that the best-performing 4D panoptic method, Eq-4D-StOP [59], is excluded because its rotation augmentation is 'detrimental to radar data [29]'. This is an unquantified design assumption, not an experimental result on RadarScenes; the cited reference concerns a different radar dataset and does not evaluate Eq-4D-StOP. Because the headline claim is comparative, the exclusion is the single most load-bearing decision in the paper. Additionally, results come from a single run without variance, and the ablation differences in Tables II and III are small (0.1-2.1 points) with no error bars, so the reported 7.5-point gain over center tracking could be partly due to evaluation protocol rather than the proposed association. A fair comparison, or a quantitative demonstration that Eq-4D-StOP cannot be adapted to radar, is required before the SOTA claim can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses moving instance tracking in sparse and noisy automotive radar point clouds. It extends the authors' Radar Instance Transformer (RIT) with two tracking-specific modules: a per-point temporal offset head that regresses each moving point to the center of its instance in the next scan, and an attention-based instance feature extractor that produces an appearance similarity cost. The final data association clusters instances with DBSCAN, performs Hungarian matching within clusters, and switches between geometric-only and combined geometric-appearance costs according to hand-set distance thresholds. Experiments on RadarScenes report LSTQ 66.8 with S_assoc 48.2, outperforming three implemented baselines, and ablations on the validation set support the contributions of the offset predictions and the attentive association.","tokens_in":12315,"tokens_out":4804,"duration_ms":52467,"significance":"If the experimental claims hold, the paper provides a sensible point-based alternative to voxelized or IoU-based trackers for radar, with a plausible mechanism for single-point instances (temporal offsets) and a modest but consistent improvement in association quality. The paper's strengths include evaluation on a public benchmark with a standard metric, implementation details sufficient for reproduction, and a set of ablations that isolate the main design choices. The main caveats are that the comparative claim rests on a selected baseline set, all numbers come from a single run, and several association thresholds are set manually without sensitivity analysis; these issues do not invalidate the method but currently temper the strength of the 'state-of-the-art' statement.","major_comments":[{"comment":"The abstract and Sec. I state that the approach achieves 'state-of-the-art performance' for moving instance tracking, but Table I compares only MOT [46], center tracking [49], and CA-Net [26]. The best-performing 4D panoptic method, Eq-4D-StOP [59], is excluded with the statement that its rotation augmentation is 'detrimental to radar data [29]'. The cited reference (View-of-Delft, Palffy et al.) does not evaluate Eq-4D-StOP, and no experiment quantifies the effect of rotation augmentation on this method. Because the headline claim is comparative, this exclusion is load-bearing. Please either add a fair adaptation of Eq-4D-StOP to the RadarScenes evaluation, provide direct evidence that it cannot be transferred, or rephrase the claim to 'improved performance over the implemented baselines'.","section":"Sec. IV-B, Table I"},{"comment":"All results are reported for a single run, with no variance or significance measures. The ablation differences in Tables II and III are small (0.1-0.3 LSTQ or S_assoc points), which is within the range of typical seed-to-seed variation for transformer-based point cloud models. The main 7.5 LSTQ gap over center tracking (Table I) is therefore not yet distinguishable from evaluation-protocol or initialization effects. Please report mean and standard deviation over at least three seeds, or provide an alternative statistical analysis.","section":"Sec. IV-A, Tables I-III"},{"comment":"The data association contains several hand-set free parameters: the DBSCAN bandwidth b=10, the geometric/appearance thresholds t_d1=5 m and t_d2=10 m, the similarity cost threshold t_c=1.5, and the track retention length of 12 scans. No sensitivity analysis is provided. These parameters directly determine when the appearance cost is used and when tracks are initialized or dropped, and their values may have been chosen with knowledge of the validation set. Please include a sensitivity study over the main thresholds (at least t_d1, t_d2, t_c, b) and state explicitly whether any parameter was tuned on the test set.","section":"Sec. III-F"},{"comment":"The temporal offset target is defined as the mean coordinate of the ground-truth instance points in the next scan. For instances consisting of a single radar detection, this target is a single noisy point whose position can jump arbitrarily between frames, and the current ablation (Table II) does not isolate this case. Since the paper motivates the center-based association partly by the single-point-instance scenario, please add an analysis or ablation that reports performance as a function of instance size (e.g., single-point vs. multi-point instances) or otherwise demonstrates stability of the temporal-offset target. Without such evidence, the claim in Sec. III-C that the temporal offset provides a regression target for single-point instances is not quantitatively supported.","section":"Sec. III-B"}],"minor_comments":[{"comment":"The notation in Eq. (7) is ambiguous: o_i is used for both offsets, and c_i is said to be 'the respective center of the instance that p_i belongs to', but for the temporal offset this must be the center in the next scan. Please define the standard and temporal offsets with distinct symbols in the loss equation.","section":"Sec. III-B, Eq. (7)"},{"comment":"The figure shows W_K and W_Q but not W_V in the attentive instance network, whereas Eq. (1) defines W_V; please align the figure with the text.","section":"Fig. 2"},{"comment":"The same workshop paper 4D-StOP appears as two separate references [20] and [21]; please merge them into a single citation.","section":"References"},{"comment":"The center tracking baseline description says Doppler velocities are 'directly used' for tracking, but the procedure is not described; please specify whether velocities are used as motion model predictions, as gating, or as association features.","section":"Sec. IV-B"},{"comment":"The row 'Geometric association t_d2=10 m' is unclear; please clarify what 'geometric association' means in this ablation, for example whether it is the same pipeline with the appearance cost removed or a pure distance threshold without the attentive similarity.","section":"Sec. IV-D, Table III"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is competently written and the method is plausible. My main hesitation is benchmark completeness: the 'state-of-the-art' claim depends on the excluded Eq-4D-StOP comparison, and the single-run results make the magnitude of the reported gain uncertain. Both issues are fixable with additional experiments, so I see major revision rather than rejection as the appropriate outcome. I would also encourage the authors to consider releasing code or a detailed appendix with per-sequence results, which would strengthen reproducibility for a journal-level contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: solid radar-specific tracking method with a clear ablation story, but the state-of-the-art claim is not yet earned. The paper does something genuinely new for radar point clouds: it predicts a temporal offset from each moving point to the center of its instance in the next scan, uses that for center-based association, and adds an attention-based appearance similarity to handle cases where geometry is ambiguous. The problem is well chosen—single-point instances break IoU-based trackers—and the evaluation on RadarScenes is a fair place to test it. The ablations support the two main design choices: adding the temporal offset improves association over the backbone, and the attentive association beats pure geometric association.\n\nThe soft spot is exactly where the stress-test note lands. Section IV-B excludes Eq-4D-StOP, the best-performing 4D panoptic method, because rotation augmentation is 'detrimental to radar data [29].' That citation is about a different dataset and does not evaluate Eq-4D-StOP. It may well be true that the augmentation hurts, but that is an assumption, not a measurement, and the paper's headline claim is explicitly comparative. Without that baseline, Table I only shows the method beats the baselines the authors chose to implement. The single-run evaluation sharpens the issue: the 7.5-point gain over center tracking could be partly protocol or seed luck, and the ablation differences of 0.1–2.1 points have no error bars. The hand-set thresholds (tc, td1, td2, DBSCAN bandwidth) also need a sensitivity analysis, since association quality can hinge on those.\n\nThe temporal-offset regression target for single-point instances is a reasonable modeling choice, not a fatal flaw, though it does assume the next-scan center is predictable from a single point.\n\nIf the comparison gap is closed, this is a useful contribution for radar-based autonomy. As is, it deserves referee time but needs revision. I'd send it out with a request for multi-seed results, a quantitative or well-argued treatment of Eq-4D-StOP, threshold sensitivity, and code release.","headline":"Solid radar tracking method, but the SOTA claim rests on an excluded baseline and single-run numbers.","tokens_in":12822,"tokens_out":2595,"would_cite":true,"duration_ms":29467,"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":"A tracker that fuses per-point temporal offsets with attention-based appearance features achieves state-of-the-art moving-instance tracking in sparse radar point clouds.","keywords":["moving instance tracking","radar point clouds","temporal offset prediction","attention-based association","center-based tracking","RadarScenes","4D panoptic segmentation","data association"],"falsifier":"Run the same tracker on radar sequences where most moving instances are single-point detections and compare the association score against a version that replaces the temporal-offset target with a velocity-based constant-motion prediction. If the association score drops to the level of the geometric-only baseline, the noisy next-scan center target is the bottleneck; alternatively, measuring track ID switches on scenes with two closely spaced pedestrians would test whether the appearance cost actually resolves geometric ambiguity.","tokens_in":11847,"feed_emoji":"📡","tokens_out":5654,"duration_ms":55677,"temperature":0.7,"pith_summary":"This paper tries to establish that moving object tracking in sparse, noisy radar point clouds can be made reliable by combining two kinds of evidence: geometric motion cues and learned appearance features. It proposes a tracker that predicts, for every radar point, a temporal offset pointing to where its instance will be in the next scan, enabling direct center-based association even when an instance is a single point. On top of that, it adds an attention-based module that extracts appearance features and computes similarity scores between tracks and new detections. On the RadarScenes benchmark, this combination reaches an LSTQ of 66.8, about 7.5 points above a center-tracking baseline, which matters because radar remains the sensor that works in bad weather and its data is exceptionally sparse.","feed_headline":"Radar tracker pairs motion centers with appearance, gains 7.5 points","feed_subtitle":"On the RadarScenes benchmark, combined association reaches an association score of 48.2 versus 38.0 for center-only tracking.","key_machinery":"The central mechanism is the per-point temporal offset prediction: for each radar point, the network regresses a vector to the center of its instance in the next scan, computed as the mean of the ground-truth points of that instance. Together with the standard offset to the current-scan center, this creates direct center-based tracks without voxelization or scan aggregation. The second mechanism is the attentive instance network, which applies local vector attention (using the Point Transformer's subtractive attention with relative positional encoding) to the moving points of each instance and aggregates them into a single instance feature vector. A similarity module then computes an attention matrix $A_{\\text{sim}} = \\text{sigmoid}(Q_{\\text{sim}}K_{\\text{sim}}^\\top + R_{\\text{sim}})$ that is converted into an appearance cost, and the data association clusters instances with DBSCAN and runs Hungarian matching locally, using the appearance cost only beyond a short-distance threshold.","core_discovery":"The paper's central claim is that a tracking-by-detection system for radar point clouds should associate instances using a cascade of geometric and appearance information rather than either alone. The geometric stage uses per-point temporal offsets to predict the next-scan center of each moving instance, making center-based Hungarian association viable in data where instances often consist of one or a few points. The appearance stage uses a transformer-based instance feature extractor and a sigmoid attention score to measure how likely two instances are the same object, which resolves the cases where geometry is ambiguous or objects interact. The final association switches from geometry-only to geometry-plus-appearance above a distance threshold, and the paper reports that this combined design outperforms the compared baselines on the RadarScenes moving-instance tracking benchmark.","pith_inferences":["If the combined association generalizes, radar-only perception stacks could provide reliable tracking in weather that disables cameras and LiDAR, since the method needs no dense point clouds or scan accumulation.","The temporal-offset target (mean of next-scan instance points) is a noisy label for single-point instances; a testable extension would be to weight the offset loss by instance point count or predict a distribution over next centers instead of a single vector.","The appearance similarity module could be trained with contrastive or metric losses to make instance features more discriminative, potentially pushing the association score further without changing the association logic.","The reported numbers are on a single benchmark with one sensor setup; a natural falsifier is running the same pipeline on another radar dataset to see whether the 7.5-point gap over center-only tracking persists across sensor configurations."],"forward_implications":["The temporal offset prediction adds motion cues that improve moving-object segmentation IoU from 84.4 to 85.4 on the validation set, and the center-based association alone gains 0.7 points in association score.","Attention-based appearance association is what closes long-distance or ambiguous matches: removing it and using only geometry at the 10 m threshold drops the association score from 54.3 to 52.2.","The method operates on single scans rather than aggregated scans, so it avoids the latency of scan accumulation, which matters for safety-critical tasks like collision avoidance.","Because the backbone is a separate moving-instance segmenter, the tracking modules can be attached to any other segmentation network, not only the one used here.","On the RadarScenes test set, the approach reaches LSTQ 66.8 with an association score of 48.2, versus 59.3 for center tracking plus Hungarian, indicating the gain is concentrated in association quality rather than classification."],"supporting_citations":[{"why":"Supplies the RadarScenes dataset and the moving-instance tracking benchmark used for training and evaluation.","marker":"[38]"},{"why":"Defines the 4D panoptic segmentation task and the LSTQ metric used to score tracking quality.","marker":"[2]"},{"why":"Provides the Radar Instance Transformer backbone that produces the segmentation, instance IDs, and point features the tracker builds on.","marker":"[51]"},{"why":"Introduces the center-based tracking baseline (CenterPoint) that the paper extends and compares against.","marker":"[49]"},{"why":"Provides the IoU-based MOT baseline and Kalman-filter tracking that demonstrate why bounding-box association fails on single-point radar instances.","marker":"[46]"},{"why":"CA-Net is the contrastive instance association baseline that also combines appearance and motion cues but with a voxel-based design.","marker":"[26]"},{"why":"Supplies the Hungarian algorithm used to solve the local bipartite assignment in the data association step.","marker":"[22]"},{"why":"Supplies the Point Transformer layer design whose subtractive attention and relative positional encoding are reused in the attentive instance network.","marker":"[57]"},{"why":"Provides DBSCAN, used to cluster instances into local areas so that Hungarian matching avoids misleading global connections.","marker":"[10]"}],"fun_headline_variants":["Radar tracker fuses motion and appearance for 7.5-point score gain","Sparse radar tracking improves when geometry and appearance are combined","Center-based radar tracking gets 7.5-point boost from appearance features","Combining motion and appearance lifts radar tracking on benchmark"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the average coordinate of the ground-truth points of an instance in the next scan is a reliable center target for every moving point, including instances that appear as a single point in the current scan.","fun_headline_variants_meta":{"raw":{"variants":["Radar tracker fuses motion and appearance for 7.5-point score gain","Sparse radar tracking improves when geometry and appearance are combined","Center-based radar tracking gets 7.5-point boost from appearance features","Combining motion and appearance lifts radar tracking on benchmark"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001173,"raw_usage":{"total_tokens":4792,"prompt_tokens":829,"completion_tokens":3963,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":3889}},"tokens_in":445,"tokens_out":3963,"duration_ms":31327,"temperature":1.0,"reasoning_tokens":3889,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:10:06.450210+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same tracker on radar sequences where most moving instances are single-point detections and compare the association score against a version that replaces the temporal-offset target with a velocity-based constant-motion prediction. If the association score drops to the level of the geometric-only baseline, the noisy next-scan center target is the bottleneck; alternatively, measuring track ID switches on scenes with two closely spaced pedestrians would test whether the appearance cost actually resolves geometric ambiguity.","supporting_citations":[{"cited_title":"Schumann, M","cited_arxiv_id":null,"evidence_quote":"Supplies the RadarScenes dataset and the moving-instance tracking benchmark used for training and evaluation."},{"cited_title":"Ayg ¨un, A","cited_arxiv_id":null,"evidence_quote":"Defines the 4D panoptic segmentation task and the LSTQ metric used to score tracking quality."},{"cited_title":"Zeller, V .S","cited_arxiv_id":null,"evidence_quote":"Provides the Radar Instance Transformer backbone that produces the segmentation, instance IDs, and point features the tracker builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the center-based tracking baseline (CenterPoint) that the paper extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the IoU-based MOT baseline and Kalman-filter tracking that demonstrate why bounding-box association fails on single-point radar instances."},{"cited_title":"Marcuzzi, L","cited_arxiv_id":null,"evidence_quote":"CA-Net is the contrastive instance association baseline that also combines appearance and motion cues but with a voxel-based design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Hungarian algorithm used to solve the local bipartite assignment in the data association step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Point Transformer layer design whose subtractive attention and relative positional encoding are reused in the attentive instance network."}],"review_version":1}