{"id":"0675676b-8426-4b65-9e0f-950c16bf4f03","arxiv_id":"2508.18506","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Radar Doppler velocities, clustered under rigidity assumptions, can be propagated to LiDAR as pseudo scene flow labels, outperforming self-supervised baselines on TruckScenes and improving label efficiency.","lead":"DoGFlow uses 4D radar Doppler measurements to estimate object motions and transfers those motion labels to LiDAR point clouds, enabling scene flow training without human annotations. On MAN TruckScenes it beats prior self-supervised methods at long range and reaches roughly 90% of fully supervised accuracy with only 10% labeled data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Doppler-only full-3D recovery in Eq. (6) is unobservable for compact clusters; pseudo-label accuracy may rest on unspecified bound constraints rather than radar evidence.","rationale":"The reader's weakest_assumption identifies the same Eq. (6) observability issue, and I agree. This is the most load-bearing point because the method's novelty is cross-modal label transfer of full 3D velocities; if velocity recovery is degenerate for typical clusters, the pseudo-labels are not physically grounded. The paper's own admission that the system 'may be over or under-constrained' (Section IV-A) plus the absence of any report of bounds or conditioning makes this a correctness risk, not merely a missing artifact. Still, the empirical comparisons (Tables I-IV, Fig. 5) are consistent with the method working in aggregate; the concern does not disprove the central claim, so the existing CONDITIONAL verdict is right. A focused observability analysis and per-cluster velocity error stratification would settle it. Secondary issues (the 2291 ms 'real-time' wording and the fact that Algorithm 1 associates only to dynamic radar points, making the majority-vote dynamic test in line 14 potentially vacuous) are worth correcting but are less central to the main claim.","tokens_in":12619,"tokens_out":7468,"duration_ms":89289,"concrete_test":"On the TruckScenes val set, for each dynamic radar cluster compute the singular value decomposition of A in Eq. (6) after ego compensation. Report (i) the fraction of clusters with condition number > 10 or effective rank < 3, (ii) the chosen bound values and the fraction of estimated velocity components that lie on a bound, and (iii) per-cluster 3D velocity error against bounding-box-derived ground truth (available every fifth frame), stratified by condition number. If high-condition clusters contribute disproportionately to EPE or have velocity estimates pinned to bounds, the full-3D label transfer claim fails; if low-condition clusters dominate the aggregate metrics, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of DoGFlow is Eq. (6): each radar point contributes one scalar equation v_j_comp = u_j^T R_ego^{Si} v_k for the 3D cluster velocity v_k. Full recovery of v_k requires the stacked row vectors u_j^T R_ego^{Si} to span R^3, i.e., sufficient angular diversity within a cluster. For a compact object at long range—the regime where DoGFlow claims its main advantage—the rows are nearly parallel, so A is rank-deficient or ill-conditioned and the least-squares solution is governed by the unstated `physical velocity bounds` [39]. The paper never reports the bound values, the fraction of rank-deficient clusters, or the conditioning of A. If many clusters are unobservable, the propagated LiDAR flow is not Doppler evidence but an unspecified regularizer, so the central claim that DoGFlow `recovers full 3D object motions` is not established. This is compounded by the ambiguity-resolution step (Algorithm 1, lines 16-19): when candidate radar-cluster velocities are already bound-dominated, choosing by Chamfer distance to Pt+1 cannot restore the missing lateral velocity component. The label-efficiency result (Fig. 5) could then be carried by Doppler-based dynamic classification plus radial motion, rather than by accurate full 3D pseudo-labels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DoGFlow proposes a self-supervised LiDAR scene flow method that generates pseudo-labels from 4D radar Doppler measurements and transfers them to LiDAR. In the first stage, dynamic radar points are thresholded by ego-compensated radial velocity, clustered via Connected Components Labeling in spatial/Doppler space, and each cluster's full 3D velocity is estimated by solving a linear system from per-point radial constraints. In the second stage, LiDAR points are denoised, clustered with HDBSCAN, associated to radar clusters, and labeled as dynamic by majority voting; ambiguous multi-velocity cases are resolved by forward projection and Chamfer distance. The authors evaluate on MAN TruckScenes, reporting improvements over self-supervised baselines in range-wise dynamic EPE and dynamic IoU, favorable pseudo-labeling results when training an SSF backbone, label-efficiency gains, and robust behavior in adverse weather.","tokens_in":12961,"tokens_out":4742,"duration_ms":58597,"significance":"The core idea is original and physically motivated: using Doppler as an independent cross-modal cue for LiDAR scene flow is a promising alternative to purely geometric self-supervision. The paper also addresses a practically important regime—long-range and adverse-weather LiDAR—where chamfer-based methods typically fail. A strength is that the label source (radar Doppler) is physically independent of the box-based ground truth used for evaluation, so the reported improvements are not trivially explained by label leakage. Another strength is the pseudo-labeling evaluation, which shows that the generated labels can train a LiDAR backbone and reduce annotation cost. However, the central claim of recovering full 3D velocities from Doppler constraints needs more rigorous support, and the empirical claims would be stronger with error bars and threshold sensitivity analysis. If these issues are addressed, the work would be a solid contribution to self-supervised scene flow.","major_comments":[{"comment":"The central velocity-recovery step treats each radar point as one scalar constraint on a 3D cluster velocity. Full recovery is only possible if the rows u_j^T R_ego^{S_i} within a cluster span R^3. For compact objects at long range—the regime where DoGFlow claims its main advantage—these rows are nearly parallel, making the system rank-deficient or ill-conditioned. The text states the system 'may be over or under-constrained' and defers to 'physical velocity bounds [39]', but the bound values, the fraction of rank-deficient clusters, and the condition numbers are never reported. If many clusters are unobservable, the propagated LiDAR flow is determined by the unspecified bounds, not by Doppler evidence. Please report observability statistics and ablations with and without the bounds.","section":"IV-A, Eq. (6)"},{"comment":"All quantitative results are single-run point estimates without error bars. For example, Table I shows DoGFlow's close-range dynamic EPE (0.6892) is worse than NSFP (0.6152) and FastNSF (0.6019), while its far-range advantage is 0.7013 vs 0.8829; without variance or significance testing it is unclear which differences are reliable. Figure 5's headline claim of 'over 90% of fully supervised performance with 10% GT' rests on a comparison of 0.4541 vs 0.4119. Please report multiple seeds with standard deviations and, if feasible, significance tests for the main comparisons.","section":"Tables I–IV and Fig. 5"},{"comment":"The method depends on at least seven manually set thresholds: δ_dyn, δ_spatial, δ_velocity, δ_intensity, δ_neighbor, and δ_adaptive^min/max. No sensitivity analysis is provided. Because the pseudo-label pipeline is not learned, these thresholds directly determine cluster boundaries and, consequently, the quality of the propagated flow. Without a sensitivity study, it is difficult to judge whether the reported gains are robust or tuned to the validation set. Please add a sensitivity sweep over the most influential thresholds, especially δ_spatial, δ_velocity, and δ_dyn.","section":"V-D and Algorithm 1"},{"comment":"The abstract and introduction claim that DoGFlow computes motion pseudo-labels 'in real-time' and enables real-time inference. However, Table I reports 2291 ± 465 ms per frame for direct scene flow estimation, which is far above the 100 ms frame interval at 10 Hz. Only the feedforward model trained on pseudo-labels runs at 25 ± 12 ms. The real-time claim for the proposed method itself is not supported. Please revise the wording or report a version of the pipeline that meets real-time constraints.","section":"Abstract and Table I"}],"minor_comments":[{"comment":"There are several typos: 'sligtly' in Section VI-A, 'leverging' in Section II-A, 'DogFlow' in the Figure 4 caption, and an extra 'and' in Section IV-A.","section":"Throughout"},{"comment":"The notation v_j^comp is used both as a scalar (Eq. (3)) and as a vector (v_j^comp = v_j^comp u_j). Please use distinct symbols to avoid ambiguity.","section":"IV-A"},{"comment":"The Doppler threshold δ_dyn = 0.05 m/s is much smaller than the dynamic GT threshold of 0.05 m per frame (equivalent to 0.5 m/s at 10 Hz). The relation between these thresholds should be discussed, since a very low Doppler threshold may classify slow or static-but-noisy returns as dynamic.","section":"V-D / VI-A"},{"comment":"The annotation schedule (bounding boxes every fifth frame) and the evaluation protocol (which frames are used for scene flow metrics) should be clarified. It is not clear whether metrics are computed only on annotated frames or on all frames.","section":"V-A"},{"comment":"The code link is only promised 'upon acceptance.' Since reproducibility is a key strength of pseudo-labeling papers, please provide an anonymous repository or more detailed implementation details, especially for the least-squares solver and bound values.","section":"Repository"}],"recommendation":"major_revision","confidential_remarks":"The paper's self-reported limitations—sensitivity to extrinsic calibration and degraded close-range Doppler—are honest, but the observability issue in Eq. (6) and the absence of any threshold sensitivity analysis need to be fixed before publication. The single-run results also make the empirical claims fragile. No concerns about the physical independence of the Doppler label source; the method is not circular with respect to the box-based ground truth."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you need to know: DoGFlow is a genuinely new cross-modal label-transfer pipeline—radar Doppler to LiDAR scene flow—and the empirical gains look real, especially at long range and in poor weather. The mechanism that recovers full 3D cluster velocities from Doppler has an observability gap the paper glosses over, plus missing reproducibility details, so read the headline claims with that in mind.\n\nWhat’s actually new: the cited work contains radar-only Doppler scene flow (Ding et al.), camera-radar full-velocity estimation (Long et al.), and multi-modal pseudo-labeling for radar (Hidden Gems), but no one has transferred Doppler-derived motion labels from 4D radar to LiDAR scene flow. That is a good idea with a sound physical basis: radial Doppler, ego-motion compensation, and rigidity are independent of the box-based ground truth and of the LiDAR input. The paper is also honest about its close-range weaknesses and sensor blind spots, which earns trust. Tables I–IV are internally coherent: DoGFlow beats the self-supervised baselines on far-range dynamic EPE and dynamic IoU, and the label-efficiency curve in Fig. 5 is the most convincing part—pretraining on the pseudo-labels gets you to within 90% of fully supervised performance at 10% labels. Table III, swapping Doppler-based dynamic classification into SeFlow, is a clean ablation.\n\nThe soft spots: the biggest is the velocity recovery in Eq. (6). Each radar point contributes one scalar equation; a compact cluster at long range yields nearly parallel line-of-sight rows, so the linear system is rank-deficient or badly conditioned. The paper says only that it uses a least-squares solver with ‘physical velocity bounds’ [39] and never reports the bound values, the conditioning, or the fraction of clusters that end up under-constrained. That is a real gap. If many clusters are bound-dominated, the propagated flow is a regularizer rather than Doppler evidence, and the label-efficiency gain may be carried mostly by dynamic classification from radial Doppler, not by accurate full-3D pseudo-labels. This is not a fatal flaw—the dynamic classification and radial components are still physically grounded—but the full-3D claim needs better support. Also: single runs without error bars, six hand-set thresholds with no sensitivity analysis, and a ‘real-time’ claim in the abstract that collides with the reported 2291 ms direct runtime (only the feedforward backbone is 25 ms). These are fixable.\n\nWho it’s for: scene flow and autonomous-driving perception researchers. It’s a solid proof-of-concept for cross-modal supervision and a fair baseline on TruckScenes.\n\nRecommendation: send it to peer review. The idea and evidence are serious enough to deserve referee time. I would ask for the conditioning report, bound values, sensitivity or error bars, and a softened runtime claim, but none of these sink the central contribution.","headline":"Real novelty and believable gains, but the Doppler-to-full-3D recovery is under-specified and needs a conditioning analysis before the full-3D claim lands.","tokens_in":13457,"tokens_out":4021,"would_cite":true,"duration_ms":46949,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DoGFlow derives dense LiDAR scene flow from radar Doppler measurements, without any manual labels.","keywords":["scene flow","LiDAR","4D radar","Doppler velocity","self-supervised learning","pseudo-labels","cross-modal label transfer","adverse weather"],"falsifier":"Select radar clusters whose points span a narrow range of line-of-sight directions at long range (e.g., a distant truck seen nearly head-on) and compare their estimated full velocity against ground truth. If DoGFlow still produces a full 3D velocity close to truth, the recovery is doing more than the Doppler constraints can support; if the velocity is pinned to the least-squares bound, the mechanism is confirmed to be rank-deficient exactly there. A simpler aggregate check is to plot cluster-level velocity error against the angular diversity of the cluster's radar points.","tokens_in":12545,"feed_emoji":"📡","tokens_out":8045,"duration_ms":85989,"temperature":0.7,"pith_summary":"The paper sets out to show that LiDAR scene flow can be learned without expensive human annotation if 4D radar Doppler measurements are used as the teacher. Because each radar point yields only a one-dimensional radial speed, DoGFlow clusters dynamic radar returns and solves for each cluster's full 3D velocity under a rigidity assumption; it then transfers those velocities to associated LiDAR clusters, using range-adaptive association and forward-projection to resolve ambiguous candidates. On the MAN TruckScenes dataset, the resulting pseudo-labels beat existing self-supervised methods, particularly beyond 35 meters and in snow, and push a LiDAR-only scene flow network to more than 90 percent of fully supervised accuracy with only 10 percent ground truth. The same pipeline can run at runtime for direct flow estimation or offline to generate training labels.","feed_headline":"Doppler radar labels give LiDAR scene flow without manual annotation","feed_subtitle":"Self-supervised DoGFlow beats prior label-free methods at long range and needs just 10% ground truth for over 90% of supervised accuracy.","key_machinery":"The load-bearing identity is the Doppler observation model v_j^comp = u_j^T R_ego^{Si} v_k, which converts each radar measurement into one scalar constraint on a three-dimensional cluster velocity. Sufficiently diverse line-of-sight directions inside a rigid cluster make the resulting linear system determine the cluster's full velocity; connected-components clustering on the spatial-plus-Doppler graph provides the rigid groups, and the ambiguity-resolved propagation step carries those velocities across to LiDAR points.","core_discovery":"The paper's central claim is that the missing motion information in LiDAR can be supplied by radar Doppler through a two-stage cross-modal transfer. First, dynamic radar points are separated by an ego-motion-compensated Doppler threshold and clustered with connected components on a spatial-plus-Doppler graph. For each cluster, the one-dimensional radial constraints from all its points form a linear system whose solution is the cluster's full 3D velocity, solved by least squares with physical bounds. Second, these velocities are propagated to LiDAR: ground-removed LiDAR is clustered by intensity, dynamic labels are assigned by majority vote over associated radar clusters, and when several can","pith_inferences":["It follows, though the paper does not test it, that the same pseudo-label recipe should transfer to other 4D-radar-equipped urban datasets, with calibration quality and radar density as the main limiting factors.","A natural stress test beyond the paper: as radar-LiDAR extrinsic calibration drifts, the range-adaptive association threshold should cause label quality to degrade smoothly; quantifying that sensitivity would show whether online extrinsic correction is needed.","Articulated objects such as pedestrians violate the per-cluster rigidity assumption; splitting clusters into locally rigid parts would extend the method but would also create more candidate velocities to disambiguate at propagation time.","The dynamic/static decision comes directly from Doppler thresholding, so the method could contribute to tracking and detection even in regimes where the exact flow vector is noisy."],"forward_implications":["LiDAR-only scene flow networks can be trained to near-supervised quality with no manual labels: on TruckScenes, an SSF backbone pretrained on DoGFlow pseudo-labels reaches about 90 percent of fully supervised performance with 10 percent ground truth.","Doppler-based dynamic classification is itself a reusable signal: swapping it into the SeFlow pipeline improves both near-range and far-range dynamic EPE and IoU.","The label-transfer pipeline is training-free and can be used either as a direct runtime estimator or as an offline pseudo-label generator, decoupling the quality of the teacher from the speed of the student.","Because radar is relatively unaffected by rain and snow, the method keeps dynamic IoU high in adverse weather where LiDAR-only geometric self-supervision degrades sharply.","The framework is modular: any LiDAR scene flow backbone can consume the pseudo-labels, so the label-efficiency result is not tied to a single network architecture."],"supporting_citations":[{"why":"Supplies the MAN TruckScenes dataset, the only evaluation benchmark combining long-range LiDAR with 360-degree 4D radar across diverse weather conditions.","marker":"[25]"},{"why":"Provides SSF, the fully supervised long-range scene flow method used as the upper-bound baseline and as the backbone for pseudo-label training and label-efficiency experiments.","marker":"[7]"},{"why":"Provides the NSFP runtime-optimization self-supervised baseline that DoGFlow must outperform in Tables I and IV.","marker":"[12]"},{"why":"Provides SeFlow, the strongest self-supervised comparison, and the pipeline whose DUFOMap dynamic classification is replaced by radar-based classification in Table III.","marker":"[13]"},{"why":"Provides FastNSF, the prior best training-free pseudo-label generator against which DoGFlow pseudo-labels are compared in Table II.","marker":"[27]"},{"why":"Provides ICP-Flow, the other leading training-free pseudo-label baseline used to train the SSF backbone in Table II.","marker":"[26]"},{"why":"Inspires the Connected Components Labeling strategy on the graph of dynamic radar points, separating nearby moving objects before velocity estimation.","marker":"[38]"},{"why":"Supplies the hierarchical clustering algorithm that groups intensity-filtered LiDAR points while excluding weather noise before label propagation.","marker":"[37]"},{"why":"Provides the bound-constrained least-squares solver that recovers cluster-level 3D velocities when the linear system from Eq. (6) is over- or under-constrained.","marker":"[39]"}],"fun_headline_variants":["Radar Doppler teaches LiDAR motion without labels","Self-supervised scene flow from radar's Doppler hints","DoGFlow: Radar guides LiDAR, no manual labels needed","10% labels, 90% accuracy: radar-guided scene flow"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes that the Doppler readings within each cluster are enough to determine all three components of the cluster's velocity; for small or distant objects that do not span many viewing angles, the estimate depends on the solver's physical bounds rather than on the measurements.","fun_headline_variants_meta":{"raw":{"variants":["Radar Doppler teaches LiDAR motion without labels","Self-supervised scene flow from radar's Doppler hints","DoGFlow: Radar guides LiDAR, no manual labels needed","10% labels, 90% accuracy: radar-guided scene flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1250,"prompt_tokens":735,"completion_tokens":515,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":447}},"tokens_in":479,"tokens_out":515,"duration_ms":5636,"temperature":1.0,"reasoning_tokens":447,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:26:16.858361+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Select radar clusters whose points span a narrow range of line-of-sight directions at long range (e.g., a distant truck seen nearly head-on) and compare their estimated full velocity against ground truth. If DoGFlow still produces a full 3D velocity close to truth, the recovery is doing more than the Doppler constraints can support; if the velocity is pinned to the least-squares bound, the mechanism is confirmed to be rank-deficient exactly there. A simpler aggregate check is to plot cluster-level velocity error against the angular diversity of the cluster's radar points.","supporting_citations":[{"cited_title":"Man truckscenes: A multimodal dataset for autonomous trucking in diverse conditions,","cited_arxiv_id":null,"evidence_quote":"Supplies the MAN TruckScenes dataset, the only evaluation benchmark combining long-range LiDAR with 360-degree 4D radar across diverse weather conditions."},{"cited_title":"Neural scene flow prior,","cited_arxiv_id":null,"evidence_quote":"Provides the NSFP runtime-optimization self-supervised baseline that DoGFlow must outperform in Tables I and IV."},{"cited_title":"SeFlow: A self-supervised scene flow method in autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Provides SeFlow, the strongest self-supervised comparison, and the pipeline whose DUFOMap dynamic classification is replaced by radar-based classification in Table III."},{"cited_title":"Fast neural scene flow,","cited_arxiv_id":null,"evidence_quote":"Provides FastNSF, the prior best training-free pseudo-label generator against which DoGFlow pseudo-labels are compared in Table II."},{"cited_title":"Icp-flow: Lidar scene flow estimation with icp,","cited_arxiv_id":null,"evidence_quote":"Provides ICP-Flow, the other leading training-free pseudo-label baseline used to train the SSF backbone in Table II."},{"cited_title":"Fully sparse 3d object detection,","cited_arxiv_id":null,"evidence_quote":"Inspires the Connected Components Labeling strategy on the graph of dynamic radar points, separating nearby moving objects before velocity estimation."},{"cited_title":"hdbscan: Hierarchical density based clustering","cited_arxiv_id":null,"evidence_quote":"Supplies the hierarchical clustering algorithm that groups intensity-filtered LiDAR points while excluding weather noise before label propagation."},{"cited_title":"A subspace, interior, and con- jugate gradient method for large-scale bound-constrained minimization problems,","cited_arxiv_id":null,"evidence_quote":"Provides the bound-constrained least-squares solver that recovers cluster-level 3D velocities when the linear system from Eq. (6) is over- or under-constrained."}],"review_version":1}