{"id":"cc559522-2fe3-4398-a699-063588c081ee","arxiv_id":"2501.11351","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An automatic LiDAR-camera labelling pipeline and a two-branch FPN plus 3D U-Net network achieve over 65% of LiDAR detection probability for semantic segmentation of 4D radar tensors on the RaDelft dataset.","lead":"This paper builds an automatic labelling pipeline for the RaDelft automotive radar dataset by fusing LiDAR and camera, then trains a 3D semantic segmentation network on 4D radar tensors using those labels. The authors report radar segmentation performance at over 65% of LiDAR detection quality, with improvements over two adapted literature baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported segmentation metrics are computed against the same automatically generated labels used as training targets; with pedestrian/bicycle F1 scores of 0.69/0.83, label noise propagates directly into Table II, so the 65.1% Pd and 0.54 m CD gains may not survive manual-label evaluation.","rationale":"I agree with the reader's weakest_assumption: the load-bearing condition is that the automatically generated labels are accurate enough to serve as ground truth, and the paper does not establish this because it evaluates the segmentation network on the same auto labels used for training. This is not merely a missing baseline; it is a circular evaluation step located in Section III.B with respect to the labels produced in Section II. The reader's additional observation that the abstract combines numbers from two configurations (65.1% Pd from Baseline VRU, 1.77 m CDAll from Baseline+Res VRU) is also valid and makes the headline claim harder to pin down, though it is secondary to the label-noise issue. I do not see an internal inconsistency in the network architecture or the automatic labelling pipeline itself; the public dataset link and the explicit F1 scores in Table I are positive signs of transparency. The proposed check on the 50 manually labelled frames is feasible because those annotations already exist in the paper, and it would settle whether the reported gains are semantic or an artifact of the label-generation process. The final recommendation remains CONDITIONAL, unchanged from the reader's verdict: the paper should be accepted only if the manual-label evaluation supports the headline numbers, or if the claims are explicitly downgraded to 'performance on auto labels'.","tokens_in":6585,"tokens_out":5696,"duration_ms":53960,"concrete_test":"Recompute Table II metrics for the Baseline (VRU) and Baseline+Res (VRU) configurations on the 50 manually labelled frames described in Section II.B, using the manual expert labels as ground truth instead of the automatic labels. If the 13.2% Pd Vehicles improvement and the 0.54 m CDAll reduction shrink, reverse, or move outside a few percentage points, the headline result is an artifact of label noise; if they persist, the central claim survives. Report per-class Pd/Pfa for pedestrians and bicycles separately, since Table I identifies these as the least reliable auto labels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests on Table II, but the evaluation protocol is circular in a way that directly affects the headline numbers. Section II.B reports the automatic labels as having F1 = 0.69 for pedestrians and 0.83 for bicycles against 50 manually labelled frames (Table I). Section III.B then uses these same auto-generated, voxelized LiDAR labels as the ground truth for computing Pd, Pfa, and Chamfer distance in Table II (the 'LiDAR ground truth' in Fig. 4 is the output of the Fig. 1 pipeline, not independent annotation). Because these labels are simultaneously the training target, a high Table II score can reflect the network learning the label-generation pipeline's systematic errors rather than true object semantics. For pedestrians, a 0.69 F1 means roughly 30% of the relevant voxel labels are wrong; this is large enough to move Pd VRU and CDTargets by more than the claimed margins. The 13.2% Pd Vehicles improvement and 0.54 m CDAll reduction compare two configurations on the same noisy labels, so the relative ordering may partially survive, but the absolute claim of 'over 65% of LiDAR detection performance' is not established for true semantic classes until the same metrics are recomputed on independent manual labels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents two contributions on the RaDelft dataset: an automatic labelling pipeline that fuses LiDAR (pre-trained Part-A2 detections, Patchwork++ ground removal, OneFormer camera calibration, and DBSCAN smoothing) and voxelizes the result into a 500×240×34 grid; and a semantic segmentation network for 4D radar tensors, consisting of two 2D FPN branches (occupancy and class latent spaces) combined and fed to a 3D U-Net. The labelling is evaluated on 50 manually labelled frames (Table I: F1 = 0.99 for scenario objects, 0.88 for vehicles, 0.83 for bicycles, 0.69 for pedestrians). The segmentation is evaluated with detection probability, false alarm rate, and Chamfer distance against the auto-generated LiDAR labels, with the best configuration at Pd All = 65.1%, Pd Vehicles = 47.9%, and CDAll = 1.77 m, compared with two literature-adapted variants and an enhanced baseline.","tokens_in":6891,"tokens_out":5232,"duration_ms":46700,"significance":"If validated, the automatic labelling pipeline would be a practical asset for radar-centric datasets lacking labels, and the public release of the generated labels on 4TU.ResearchData is a genuinely useful contribution. The proposed network is also among the first attempts at direct voxel-level semantic segmentation from range-azimuth-elevation-Doppler radar tensors. However, the central quantitative claim about segmentation quality is undercut by a circular evaluation: the auto-generated labels serve as both training targets and test ground truth, and the labelling step's own pedestrian F1 of 0.69 shows substantial label noise. Hence the significance of the segmentation results cannot be assessed from the reported numbers alone.","major_comments":[{"comment":"The segmentation evaluation is circular: the 'LiDAR ground truth' in Fig. 4 is the output of the automatic labelling pipeline of Fig. 1, and these same auto-generated voxel labels are used as the training targets for the network (Section III.A). Table I shows the auto labels have only 0.69 F1 for pedestrians and 0.83 for bicycles against 50 manual frames, so label noise propagates directly into every metric in Table II. A network that learned the label-generation pipeline's systematic errors would score higher on these metrics without being semantically better. Since the 50 manually labelled frames are used only for labelling evaluation and not for scoring the segmentation, the claimed 'over 65% of LiDAR detection performance' is not established against independent ground truth. Please recompute Table II on the manually labelled frames (or on a held-out clean subset) and report both sets of numbers.","section":"III.B / Table II and Fig. 4"},{"comment":"The headline numbers in the abstract and conclusion are not reproducible from Table II. The '13.2% improvement in vehicle detection probability' is the difference between Baseline (VRU) at 47.9% and Baseline at 34.8%, i.e., the effect of merging pedestrian and bicycle classes within the authors' own baseline, not a comparison against the literature variants. The '0.54 m reduction in Chamfer distance' is the difference between Variant 1 (2.31 m) and Baseline + Res (VRU) (1.77 m) in CDAll, which compares a single variant to a different proposed configuration. Also, the best Pd All (65.1%) comes from Baseline (VRU), while the best CDAll (1.77 m) comes from Baseline + Res (VRU); no single configuration attains all the abstract's claims. Please state exactly which configuration each reported number refers to and use a consistent comparison.","section":"Abstract and Section IV"},{"comment":"The phrase 'over 65% of the LiDAR detection performance' is undefined as stated. Pd All is the detection probability of radar voxels with respect to the LiDAR-derived labels; it is not a ratio to LiDAR's own detection performance. Please either define the quantity precisely or rephrase the claim.","section":"III.B.2"}],"minor_comments":[{"comment":"The heading 'Tranformation and voxelization' and the figure label 'Tranformation' contain a typo: it should be 'Transformation'.","section":"Fig. 1 and Section II.A.3"},{"comment":"Training details such as learning rate, batch size, number of epochs, and the weights of the wCE and SDice losses are not reported; these are needed for reproducibility.","section":"III.A"},{"comment":"Since the test set comprises only two scenes and each configuration is run once, the absence of error bars or per-frame variance makes it difficult to judge whether the differences between configurations are meaningful; adding such information would strengthen the comparison.","section":"Table II"},{"comment":"The calibration distance threshold (25 m) and DBSCAN parameters (epsilon = 0.6, minPts = 100) are presented without sensitivity analysis; a brief discussion of how these choices affect the labelling quality would help.","section":"II.A.2"},{"comment":"The CDTargets metric is said to be high because static and dynamic targets are not distinguished; please clarify how 'scenario objects' relates to parked vehicles and whether the high CDTargets is partly an artifact of the label definition.","section":"III.B.2"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be a condensed version of the first author's MSc thesis, and the public release of the labels is a positive step. The main barrier is the circular evaluation; if the authors can recompute the segmentation metrics on the 50 manually labelled frames, the central claim could become credible. The abstract's aggregated numbers also need to be aligned with the actual configurations in Table II. This is not a reject, but the current version overstates what the experiments establish."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper builds a plausible auto-labelling pipeline for 4D radar and shows a segmentation network on top, but the headline numbers are compromised because Table II is computed against the very auto-labels used for training. That makes the absolute claims untrustworthy, even though the relative ordering between methods may partly hold.\n\nWhat's actually new: an automatic labelling process for the RaDelft dataset, combining Part-A2 detections from LiDAR, OneFormer semantic calibration from camera, and DBSCAN smoothing, with a public release of the generated labels. That is a useful contribution to the radar perception community. The two-branch FPN plus 3D U-Net on RAE tensors is a reasonable design; nothing groundbreaking, but it's a workable baseline. The manual labelling of 50 frames for the labelling evaluation is also a good effort.\n\nSoft spots, in order: The main issue is the self-referential evaluation. Table II uses voxelized LiDAR auto-labels as ground truth, and those same labels are the training target for the segmentation network. With pedestrian/bicycle F1 around 0.69/0.83, the label noise is non-trivial and directly inflates the reported Pd and CD numbers. The fix is straightforward: re-evaluate the final model on the 50 manually labelled frames (or a larger manual set) and report class-wise metrics there. Without that, the 'over 65% of LiDAR detection' claim is not established. Second, the abstract and conclusion mix numbers from different configurations: the 65.1% Pd All and 47.9% Pd Vehicles come from 'Baseline (VRU)', while the 1.77m CDAll comes from 'Baseline + Res (VRU)'. This is sloppy and should be cleaned up. Third, the literature baselines are adapted in a somewhat ad hoc way; the comparison is acceptable for a workshop paper but not a strong benchmark.\n\nWho this is for: people working on radar semantic segmentation and multi-sensor fusion for ADAS. It's a solid empirical paper with a fixable methodological gap. I think it deserves a serious referee, and with the manual-label re-evaluation it could become a useful reference. I would cite the dataset and labelling approach, not the segmentation numbers.","headline":"Useful auto-labelling contribution for 4D radar, but the segmentation evaluation is circular and the headline numbers mix configurations.","tokens_in":7377,"tokens_out":2402,"would_cite":true,"duration_ms":21726,"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":"This paper claims that 4D radar tensors alone can be semantically segmented into scenario objects, vehicles, and vulnerable road users, using automatically generated labels from LiDAR and camera, reaching over 65% of LiDAR-based detection…","keywords":["4D automotive radar","semantic segmentation","automatic labelling","LiDAR-camera fusion","RAED radar tensor","RaDelft dataset","voxel classification","radar point cloud"],"falsifier":"Manually label the two unseen test scenes (or reuse the 50-frame manual set) and recompute detection probability and Chamfer distance against those manual labels; if the numbers drop materially below 65.1%, 47.9%, and 1.77 m, the reported segmentation performance is an artefact of auto-label error propagation rather than a radar capability.","tokens_in":6407,"feed_emoji":"📡","tokens_out":4853,"duration_ms":43892,"temperature":0.7,"pith_summary":"This paper tackles two gaps at once: automotive radar lacks labelled data, and 4D radar tensors are rarely used for semantic segmentation. It proposes an automatic labelling pipeline that combines a LiDAR-only 3D detector, camera semantic segmentation, and density-based clustering to create point-by-point labels for the RaDelft dataset. Those labels train a voxel-wise segmentation network that maps a range-azimuth-elevation-Doppler radar tensor to class probabilities per voxel. The central claim is that radar alone, with automatically generated labels, can recover over 65% of the LiDAR detection performance and nearly 50% of vehicle detection, while beating two literature-inspired variants on detection probability and Chamfer distance.","feed_headline":"4D radar segmentation reaches 65% of LiDAR detection","feed_subtitle":"Auto-generated labels from camera and LiDAR train a voxel classifier that beats literature-inspired radar baselines.","key_machinery":"The load-bearing mechanism is a two-stage pipeline. First, automatic labelling: a pretrained LiDAR detector (Part-A2) proposes bounding boxes, Patchwork++ removes ground points, the OneFormer camera model re-labels close points within 25 m, and DBSCAN enforces label consistency by majority voting; the labels are then voxelized into a 500 x 240 x 34 range-azimuth-elevation cube. Second, segmentation: the radar tensor is projected to the same cube, a 2D FPN with ResNet-18 backbone produces an occupancy latent space and a class latent space, broadcasting merges them into a 3D semantic latent space, and a five-layer 3D U-Net outputs per-voxel class probabilities. The paper's numbers stand or fall on this division of labour: the labelling must be trustworthy enough to serve as ground truth, and the voxelized radar power must preserve the class-discriminative information.","core_discovery":"The central discovery is that a 4D radar tensor, converted to a range-azimuth-elevation power cube, contains enough information for voxel-level semantic segmentation when paired with labels auto-generated from LiDAR and camera. The proposed network uses a two-branch 2D backbone to build occupancy and class latent spaces, combines them by broadcasting, and refines the result with a 3D U-Net. On the RaDelft test scenes, the best configuration (with pedestrians and bicycles merged into a vulnerable-road-user class) achieves 65.1% detection probability for all points, 47.9% for vehicles, and 1.77 m Chamfer distance, outperforming the compared variants by 13.2 percentage points in vehicle detection and by 0.54 m in Chamfer distance.","pith_inferences":["Because the evaluation metrics in Table II are computed against the same automatically generated labels used for training, the reported 65.1% detection probability is best read as an upper bound; an evaluation against manual labels on the same test frames would likely be lower, especially for pedestrians, whose auto labels have only 0.69 F1-score.","The labelling pipeline could transfer to other radar datasets if LiDAR and camera extrinsics are available, but the 25 m camera calibration range and the KITTI-trained detector will limit label quality for small, distant, or closely packed objects.","Doppler is discarded when the RAED tensor is averaged into the RAE cube; keeping Doppler as an extra channel could plausibly improve VRU detection, since the paper's own qualitative result credits Doppler for dynamic vehicles.","The comparison against literature variants is indirect because those variants were adapted from detection networks; a direct comparison on a common labelled benchmark would clarify whether the gain comes from the two-branch architecture or from the training labels."],"forward_implications":["Semantic segmentation of 4D radar is feasible without manual annotation, since the automatic labelling pipeline replaces human labelling for the RaDelft dataset.","Merging pedestrians and bicycles into a vulnerable-road-user class improves detection probability for all points and reduces Chamfer distance, so class granularity trades against segmentation accuracy.","Radar's Doppler information helps segment dynamic targets, with moving vehicles among the best-segmented objects in the qualitative results.","The publicly released dataset with generated labels enables future radar-only perception work that does not need LiDAR at inference time."],"supporting_citations":[{"why":"Supplies the RaDelft dataset, including 4D radar tensors, LiDAR point clouds, RGB images, and collection setup used throughout.","marker":"[6]"},{"why":"Provides the pretrained Part-A2 LiDAR detector that generates preliminary bounding-box labels.","marker":"[8]"},{"why":"Supplies the KITTI dataset on which the pretrained detector was trained.","marker":"[9]"},{"why":"Patchwork++ ground segmentation is used to remove ground points before calibration.","marker":"[10]"},{"why":"OneFormer camera semantic segmentation model is used to calibrate labels within 25 m.","marker":"[11]"},{"why":"DBSCAN clustering is used for label consistency adjustment by majority voting.","marker":"[13]"},{"why":"Feature Pyramid Network structure is used as the 2D backbone for both latent spaces.","marker":"[14]"},{"why":"3D U-Net architecture is used as the 3D backbone for final segmentation.","marker":"[15]"},{"why":"Variant 1 comparison: a data-driven radar detector adapted into a segmentation network.","marker":"[16]"},{"why":"Variant 2 comparison: K-Radar detector adapted with sparse input, used as a baseline.","marker":"[17]"}],"fun_headline_variants":["4D radar segmentation hits 65% of LiDAR performance","Auto-labeled radar achieves 65% of LiDAR detection","Radar semantic segmentation reaches 65% LiDAR level","4D radar + auto labels: 65% of LiDAR accuracy","4D radar with auto labels matches 65% of LiDAR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The automatically generated labels are accurate enough to serve as ground truth for training and evaluating the segmentation network, despite F1-scores of only 0.69 for pedestrians and 0.83 for bicycles on 50 manually labelled frames.","fun_headline_variants_meta":{"raw":{"variants":["4D radar segmentation hits 65% of LiDAR performance","Auto-labeled radar achieves 65% of LiDAR detection","Radar semantic segmentation reaches 65% LiDAR level","4D radar + auto labels: 65% of LiDAR accuracy","4D radar with auto labels matches 65% of LiDAR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000502,"raw_usage":{"total_tokens":2385,"prompt_tokens":805,"completion_tokens":1580,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":1493}},"tokens_in":421,"tokens_out":1580,"duration_ms":13049,"temperature":1.0,"reasoning_tokens":1493,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:21:25.704946+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Manually label the two unseen test scenes (or reuse the 50-frame manual set) and recompute detection probability and Chamfer distance against those manual labels; if the numbers drop materially below 65.1%, 47.9%, and 1.77 m, the reported segmentation performance is an artefact of auto-label error propagation rather than a radar capability.","supporting_citations":[{"cited_title":"A deep automotive radar detector using the radelft dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the RaDelft dataset, including 4D radar tensors, LiDAR point clouds, RGB images, and collection setup used throughout."},{"cited_title":"From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network,","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained Part-A2 LiDAR detector that generates preliminary bounding-box labels."},{"cited_title":"Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3D point cloud,","cited_arxiv_id":null,"evidence_quote":"Patchwork++ ground segmentation is used to remove ground points before calibration."},{"cited_title":"Oneformer: One transformer to rule universal image segmentation,","cited_arxiv_id":null,"evidence_quote":"OneFormer camera semantic segmentation model is used to calibrate labels within 25 m."},{"cited_title":"3d u-net: learning dense volumetric segmentation from sparse annotation,","cited_arxiv_id":null,"evidence_quote":"3D U-Net architecture is used as the 3D backbone for final segmentation."},{"cited_title":"See further than cfar: a data-driven radar detector trained by lidar,","cited_arxiv_id":null,"evidence_quote":"Variant 1 comparison: a data-driven radar detector adapted into a segmentation network."}],"review_version":1}