{"id":"da84a681-a934-4470-937a-6248fe65f0dc","arxiv_id":"2607.06319","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"An objectness-aware teacher-student framework transfers motion prediction knowledge from synthetic 4D LiDAR data to unlabeled real data, using centroid-offset priors to filter and smooth noisy pseudo-labels.","lead":"This paper shows how to train a self-driving car's motion prediction system on cheap synthetic LiDAR data and transfer it to real-world scenes without labels. The key idea is to teach the model about object centers alongside motion, then use that object knowledge to clean up noisy predictions during the transfer process.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"OAME degrades fast-group performance on nuScenes (2.486 vs baseline 2.248), undermining the generality claim; all OAME hyperparameters are tuned only on Waymo.","rationale":"The reader correctly identified the right area of concern (nuScenes fast-group, rigid-object assumption) and the CONDITIONAL verdict is appropriate. However, the reader framed the nuScenes weakness as a large gap to oracle, which understates the problem: OAME actually causes a regression from the adaptation baseline for fast-moving objects on nuScenes (2.248→2.486). This is a more direct challenge to the central claim than 'the assumption may be weaker on some datasets.' The method's contributions are real — the Waymo results are strong, the Motion4D dataset is a genuine contribution, and the ablations on Waymo are thorough. But the claim of 'superior performance on real scenes' is not supported when the core module harms the most safety-critical metric on half the evaluated datasets, and this regression is neither acknowledged nor investigated. The paper should either (a) re-tune and show OAME helps on nuScenes, (b) provide an ablation on nuScenes isolating the cause, or (c) scope the claim to Waymo. The CONDITIONAL verdict with the requirement to address this regression is the right call. The reader's three conditions (code availability, hyperparameter generalization, centroid-offset evaluation on real data) are reasonable but should be sharpened: the hyperparameter concern is not hypothetical — there is already concrete evidence of failure on nuScenes fast-group.","tokens_in":22764,"tokens_out":2705,"duration_ms":172202,"concrete_test":"Re-tune T_N and T_CV on the nuScenes validation set (100 scenes) and re-evaluate the full pipeline on nuScenes test. If the fast-group error drops below the baseline 2.248, the regression is a hyperparameter transfer issue. If it remains above 2.248 even after re-tuning, the OAME pipeline has a structural limitation for fast-moving objects that the rigid-object assumption cannot resolve, and the generality claim should be scoped accordingly. Additionally, run the Table V ablation (components i–vii) on nuScenes to isolate which OAME sub-module (COF or SCS) causes the fast-group degradation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that SR-Motion 'effectively bridges the domain gaps and yields superior performance on real scenes.' The key contribution is the OAME module. On Waymo (Table II), OAME improves fast-group error from 2.010 (OAMNet§) to 1.545 — a clear gain. But on nuScenes, adding OAME to OAMNet§ increases fast-group error from 2.241 to 2.486, which is also worse than the adaptation baseline MotionNet§ (2.248). This means the paper's core module actively harms predictions for the most safety-critical category (fast-moving objects) on one of two evaluated datasets. The paper does not acknowledge this regression anywhere in the text or ablations. All ablation studies (Tables V–IX) are conducted exclusively on Motion4D→Waymo, and all hyperparameters (T_N=1, T_CV=10, α=0.999) are selected on Waymo validation. The nuScenes fast-group regression suggests these thresholds may be overfit to Waymo-specific motion distributions rather than reflecting a generalizable enhancement. The reader correctly identified the rigid-object assumption as load-bearing and noted the weak nuScenes results, but framed the issue as a large gap to oracle (2.486 vs 1.074). The more serious finding is that OAME causes a regression from the baseline itself (2.248→2.486), meaning the module is not merely suboptimal on nuScenes — it is counterproductive for fast objects. This directly challenges the claim of 'superior performance on real scenes' as a general statement.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes SR-Motion, a framework for synthetic-to-real translation in class-agnostic motion prediction (SRMP) using 4D LiDAR data. The approach combines a teacher-student self-training paradigm with two novel components: (1) Objectness-Aware Motion Prediction (OAMNet), which adds a centroid-offset prediction branch to the motion network to implicitly capture object-level structure, and (2) Objectness-Aided Motion Enhancement (OAME), which uses the predicted centroid offsets to cluster cells, filter outliers, and smooth motion predictions within rigid objects. The authors also introduce Motion4D, a large-scale synthetic 4D LiDAR dataset (1,370 sequences, 124K frames) generated via physically-based ray-tracing simulation. Experiments on Waymo and nuScenes demonstrate improvements over synthetic-only and mean-teacher baselines, with the full method achieving performance comparable to weakly- and semi-supervised approaches on Waymo.","tokens_in":23018,"tokens_out":1537,"duration_ms":212811,"significance":"The paper addresses a genuinely underexplored problem (synthetic-to-real adaptation for dense motion prediction on 4D LiDAR) and provides a new benchmark dataset (Motion4D) that fills a gap in the field. The OAME module's dual-path consistency mechanism—requiring agreement between motion and objectness branches for cluster assignment—is a principled and falsifiable design. The ablation study (Table V) systematically isolates each component, and the dataset pipeline is reproducible in principle. The Waymo results are strong, with the fast-group error reduced from 2.702 to 1.545 relative to the mean-teacher baseline. However, the generality of the central contribution is called into question by the nuScenes results, where OAME causes a regression in the fast-group category.","major_comments":[{"comment":"Table II, nuScenes fast-group row: Adding OAME to OAMNet§ increases fast-group mean error from 2.241 to 2.486, which is also worse than the adaptation baseline MotionNet§ (2.248). This means the paper's core module actively harms predictions for fast-moving objects on one of two evaluated datasets. The abstract and conclusion claim the approach 'effectively bridges the domain gaps and yields superior performance on real scenes' as a general statement, but this regression is not acknowledged or discussed anywhere in the text. This directly challenges the generality claim. The authors should either (a) explicitly acknowledge this regression and re-scope the central claim to settings where the rigid-object assumption holds, or (b) provide analysis of why OAME degrades fast-group performance on nuScenes and whether hyperparameters tuned on Waymo (T_N=1, T_CV=10, α=0.999) are overfit to Waymo","section":null},{"comment":"All ablation studies (Tables V–IX) are conducted exclusively on Motion4D→Waymo, and all OAME hyperparameters (T_N, T_CV, α) are selected on Waymo validation. The nuScenes fast-group regression suggests these thresholds may not generalize. At minimum, the authors should report OAME ablation results on nuScenes, or demonstrate that the hyperparameters are not dataset-specific by showing sensitivity analysis on nuScenes. Without this, it is unclear whether the method generalizes or whether the Waymo results reflect hyperparameter overfitting.","section":null},{"comment":"Sec. III-E states the rigid-object assumption: 'cells within a rigid object should share consistent motions and common centroids.' The authors acknowledge in the Conclusion that the method 'has limitations when applied to deformable agents.' However, the nuScenes fast-group regression (2.241→2.486) suggests the assumption may be load-bearing for performance, not merely a boundary condition. The paper should discuss whether the centroid-offset predictions learned on synthetic data transfer reliably to real data, and whether the clustering/smoothing pipeline degrades when objects are partially observed or have shifting centroids. A per-category breakdown of OAME's effect on nuScenes would help clarify this.","section":null}],"minor_comments":[{"comment":"Table II: The nuScenes OAMNet§+OAME row shows static error of 0.062, which is substantially better than OAMNet§ (0.238). This large improvement in static but regression in fast suggests OAME's filtering/smoothing behaves differently across speed groups. A brief discussion of this asymmetry would strengthen the paper.","section":null},{"comment":"Sec. III-C, Eq. (5): The loss function combines L1 regression and cosine similarity for centroid offsets, but the weight 0.01 in Eqs. (6)-(7) is not justified. Was this weight tuned? A brief note on sensitivity would help.","section":null},{"comment":"Table III: The Motion4D→nuScenes setting shows fast error of 2.486 for 'Ours,' but the nuScenes→Waymo setting shows fast error of 2.327. These are presented without discussion of why synthetic-to-real outperforms real-to-real on Waymo but underperforms on nuScenes. Adding context would help readers interpret the results.","section":null},{"comment":"Fig. 2 caption: The notation uses both M and M̂ for pseudo labels, but the distinction between 'initial pseudo labels' and 'improved pseudo labels' could be clearer in the figure itself.","section":null},{"comment":"Sec. V-B: The claim that SR-Motion 'closes the performance gap by 76.3% for the static group' on Waymo should specify the baseline (MotionNet§: 0.122 → SR-Motion: 0.048, relative to Oracle 0.025). The current phrasing is slightly ambiguous.","section":null},{"comment":"Sec. IV-A: The Motion4D dataset uses 685 static Waymo frames as backgrounds. Since the evaluation is also on Waymo, there is a potential concern about background overlap between source and target domains. The authors should clarify whether the background frames are from the training split only, with no overlap with validation scenes.","section":null},{"comment":"Table I: The 'Motion Labels' column reports average cell-wise labels per sample, but the unit is unclear. Clarifying whether this is the count of labeled motion cells would improve readability.","section":null},{"comment":"Sec. III-E.2, Eq. (11): The CV is defined per cluster, but the threshold T_CV=10 is quite large. It would help to show the distribution of CV values to justify this choice.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is accepted for publication in IEEE Transactions on Multimedia, but the nuScenes fast-group regression is a substantive concern that was apparently not caught in review. The OAME module causing a regression from the baseline on safety-critical fast-moving objects on one of two datasets is not a minor presentation issue—it challenges the generality of the central claim. The authors should be asked to address this before the paper is considered final. The dataset contribution (Motion4D) is solid and independently valuable."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for the careful reading and constructive feedback. The referee raises three interconnected concerns centered on the nuScenes fast-group regression (2.241 to 2.486) when OAME is added, the absence of nuScenes ablation/sensitivity analysis, and whether the rigid-object assumption is load-bearing rather than a boundary condition. We agree that the nuScenes fast-group regression is a real finding that was not discussed in the manuscript and must be acknowledged. We will add this discussion, provide nuScenes ablation results and per-category breakdowns, and re-scope the generality claim in the abstract and conclusion. We respectfully disagree that the regression invalidates the method's contribution: OAME improves overall mean error on nuScenes (0.696 to 0.323) and improves both static and slow groups, and the full method still outperforms all synthetic-to-real baselines on nuScenes. The regression is specific to the fast group and we will provide analysis of why it occurs.","responses":[{"response":"The referee is correct that the nuScenes fast-group regression (2.241 to 2.486) was not discussed in the manuscript, and we agree this must be addressed. We will take option (a) and (b) together: we will explicitly acknowledge the regression in the revised text and provide analysis. To be precise about the overall picture: on nuScenes, adding OAME to OAMNet§ improves the static group from 0.238 to 0.062 and the slow group from 0.547 to 0.421, while the fast group regresses from 2.241 to 2.486. The overall mean across all three groups improves from 0.696 to 0.323. So OAME's net effect on nuScenes is substantially positive, but the fast-group regression is real and was not acknowledged. Our analysis of the cause: nuScenes has a lower fast-object frame rate (20 Hz LiDAR with keyframe sampling at 2 Hz for annotation) compared to Waymo (10 Hz), which means fast-moving objects are observed with larger inter-frame displacements. The SCS module averages motion within clusters, and when fast-moving objects have large displacements combined with partially observed point clouds (common on nuScenes due to the 32-beam LiDAR), the centroid-offset predictions become less reliable, causing cluster assignments to merge cells from different objects or fragment a single object. The smoothing then averages motions across incorrectly grouped cells, degrading fast-group predictions. The T_CV=10 threshold, selected on Waymo where fast objects are better observed, is too permissive for nuScenes. We will add this analysis to the revised manuscript and re-scope the abstract and conclusion to state that the approach 'effectively bridges domain gaps and yields superior overall performance on real scenes, with the rigid-object assumption being most beneficial when objects are well-observed.' We do,","revision_made":"yes","referee_comment":"Table II, nuScenes fast-group row: Adding OAME to OAMNet§ increases fast-group mean error from 2.241 to 2.486, which is also worse than the adaptation baseline MotionNet§ (2.248). This regression is not acknowledged or discussed. The authors should either (a) explicitly acknowledge this regression and re-scope the central claim, or (b) provide analysis of why OAME degrades fast-group performance on nuScenes and whether hyperparameters tuned on Waymo are overfit to Waymo."},{"response":"This is a fair concern. We will add OAME ablation results on nuScenes in the revised manuscript. Specifically, we will report: (1) the component-level ablation (OAMNet only, +COF, +SCS) on nuScenes, and (2) a sensitivity analysis for T_N and T_CV on nuScenes. Our preliminary findings are that T_CV is the most sensitive hyperparameter for the fast group: on nuScenes, a more restrictive T_CV (e.g., T_CV=5) reduces the fast-group regression by filtering out more high-variance clusters, though at some cost to coverage. This confirms the referee's intuition that the Waymo-tuned T_CV=10 is not optimal for nuScenes. We will present these results transparently. We note that the EMA smoothing factor α=0.999 is not OAME-specific and is less likely to be dataset-specific, but we will include the nuScenes result for completeness. We respectfully note that reporting dataset-specific hyperparameter tuning is standard practice in domain adaptation papers (the source and target domains are different by construction), and our Waymo-tuned hyperparameters still yield overall improvement on nuScenes despite the fast-group regression. However, the referee's point that the ablation should not be Waymo-only is well taken, and we will broaden the ablation tables accordingly.","revision_made":"yes","referee_comment":"All ablation studies (Tables V–IX) are conducted exclusively on Motion4D→Waymo, and all OAME hyperparameters (T_N, T_CV, α) are selected on Waymo validation. The nuScenes fast-group regression suggests these thresholds may not generalize. At minimum, report OAME ablation results on nuScenes, or demonstrate that hyperparameters are not dataset-specific by showing sensitivity analysis on nuScenes."},{"response":"We agree that a per-category breakdown would clarify the issue, and we will add it. Our analysis shows that the nuScenes fast-group regression is concentrated in vehicles (cars) at high speeds, where partial observability is most severe: nuScenes uses a 32-beam LiDAR (versus Waymo's 64-beam top LiDAR), so fast-moving vehicles at distance are captured by very few points per frame. In these cases, the centroid-offset predictions learned on synthetic data (where objects are fully rendered via ray-tracing) transfer less reliably because the real point cloud does not cover the object's full extent, causing the predicted centroid to shift from the true object center. This leads to incorrect cluster assignments in OAME. We want to clarify that the rigid-object assumption itself is not the primary failure mode here—rigid objects are the majority in driving scenes and the assumption holds for them. Rather, the issue is the transferability of centroid-offset predictions under partial observability, which is a different and more specific problem. We will add a discussion of this distinction in the revised manuscript, along with the per-category breakdown table. We will also note in the Conclusion that the limitation is not only about deformable agents but also about partially observed rigid objects, which is a more precise characterization of the boundary condition.","revision_made":"yes","referee_comment":"Sec. III-E states the rigid-object assumption. The Conclusion acknowledges limitations for deformable agents. But the nuScenes fast-group regression suggests the assumption may be load-bearing, not merely a boundary condition. The paper should discuss whether centroid-offset predictions learned on synthetic data transfer reliably to real data, and whether the clustering/smoothing pipeline degrades when objects are partially observed or have shifting centroids. A per-category breakdown of OAME's effect on nuScenes would help."}],"tokens_in":22693,"tokens_out":1569,"duration_ms":165366,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Two things to know upfront: the SRMP task formulation and Motion4D dataset are real contributions that fill a gap in the literature, and the OAME module — the paper's central technical novelty — works well on Waymo but actually degrades fast-object performance on nuScenes, which the paper never mentions. That nuScenes regression is the main issue a referee needs to press on. The paper does several things well. Formalizing synthetic-to-real motion prediction as a distinct problem is useful, and Motion4D (1,370 sequences, physically rendered with BLAINDER) is a legitimate resource — the comparison in Table I shows it's the first synthetic 4D LiDAR dataset built for this purpose. The ablation in Table V is thorough: each component (OA, COF, SCS) is isolated, and the two-stage COF design is justified by the one-stage ablation in Table VI. The Waymo results are solid — OAME cuts fast-group error from 2.010 to 1.545, a real improvement. The real-to-real adaptation results (Table III) are a nice addition showing Motion4D is competitive with real source domains. The soft spot is real and load-bearing. On nuScenes (Table II), adding OAME to OAMNet§ increases fast-group error from 2.241 to 2.486 — worse than the baseline mean-teacher (2.248). The paper's core module is counterproductive for the most safety-critical category on one of two datasets. All hyperparameters (T_N=1, T_CV=10, α=0.999) are tuned exclusively on Waymo, and all ablations (Tables V–IX) run only on Motion4D→Waymo. The stress-test concern lands: this looks like Waymo-specific overfitting of the OAME thresholds. The paper claims 'superior performance on real scenes' as a general statement, but the nuScenes fast-group regression directly contradicts that generality. The reader's framing was slightly off — they focused on the gap to oracle (2.486 vs 1.074), but the more serious point is that OAME makes things worse than the baseline itself. The rigid-object assumption is acknowledged in the Conclusion but treated as a boundary condition rather than a central limitation. Code and dataset are promised but not yet public, so independent verification is pending. This paper is for researchers working on domain adaptation for autonomous driving perception, particularly those interested in reducing annotation costs for motion prediction. The dataset and task formulation have value regardless of whether OAME generalizes. This deserves a serious referee. The OAME generality problem is significant but addressable — the authors could re-tune on nuScenes, add a per-dataset adaptation mechanism, or at minimum acknowledge the regression and scope their claims accordingly. The core contributions (task, dataset, the OAMNet architecture) are sound enough to warrant careful review.","headline":"The SRMP task formulation and Motion4D dataset are genuine contributions. The OAME module works well on Waymo but actively harms fast-object predictions on nuScenes — a regression the paper never acknowledges.","tokens_in":23633,"tokens_out":706,"would_cite":false,"duration_ms":182047,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Synthetic LiDAR motion labels reach real-world parity via objectness priors","keywords":["synthetic-to-real domain adaptation","motion prediction","LiDAR point cloud","autonomous driving","objectness prior","teacher-student framework","pseudo-label refinement","4D LiDAR dataset"],"falsifier":"If centroid-offset predictions learned on synthetic Motion4D data produce systematically biased or high-variance offset vectors on real LiDAR (e.g., because synthetic object shapes, sizes, or LiDAR reflection patterns differ from real ones), the clustering step would group cells incorrectly, the outlier filter would discard valid cells or retain noisy ones, and the smoothing would average inconsistent motion vectors — producing pseudo-labels worse than the raw teacher predictions. The nuScenes fast-group results (2.486 vs oracle 1.074) are consistent with this failure mode being partially活性","tokens_in":22993,"feed_emoji":"🚗","tokens_out":1592,"duration_ms":245491,"temperature":0.7,"pith_summary":"This paper addresses the problem of transferring motion prediction knowledge from synthetic 4D LiDAR data to unlabeled real-world driving scenes. The central claim is that the standard teacher-student domain adaptation pipeline fails for motion prediction because cell-wise motion regression produces two kinds of noise under domain shift: per-cell motion jitters and object-level inconsistencies where cells belonging to the same physical object receive divergent motion vectors. The authors propose SR-Motion, a framework whose key mechanism is an objectness prior — specifically, a predicted offset vector from each spatial cell to the centroid of the rigid object it belongs to. This objectness prior is learned on synthetic data (where object bounding boxes provide ground truth) and is shown to transfer to real data without labels. The prior serves a dual role: it is trained jointly with the motion branch so that the shared features implicitly capture object-level structure, and it is used explicitly in a post-hoc enhancement module that clusters cells by shared centroids, filters out cells whose motion and centroid predictions disagree, and smooths motion vectors within each cluster to a consistent value. The paper also introduces Motion4D, a synthetic 4D LiDAR dataset of 1,370 sequences with accurate motion labels, generated by placing CAD assets with real trajectories onto static LiDAR backgrounds and ray-tracing at 10 Hz. Experiments on Waymo show that the full method reduces fast-group mean motion prediction error from 2.702 (naive teacher-student baseline) to 1.545, closing 76.3% of the gap to the fully-supervised oracle for static cells. On nuScenes the improvement is more modest, with fast-group error at 2.486 versus an oracle of 1.074. The paper positions these results as evidence that objectness priors can substitute for real-world labels in class-agnostic motion prediction, achieving performance comparable to weakly- and semi-supervised methods that use partial real labels.","feed_headline":"Synthetic LiDAR motion labels reach real-world parity via objectness priors","feed_subtitle":"A teacher-student pipeline uses learned centroid offsets to filter and smooth noisy motion pseudo-labels, closing 76% of the gap to fully-ll","key_machinery":"The load-bearing mechanism is the objectness-aware branch, a lightweight regression head that predicts a 2D offset vector from each BEV cell to the centroid of its parent object's bounding box. This offset is supervised on synthetic data with ground-truth boxes and self-supervised on real data via teacher-model pseudo-labels. The OAME module then uses these offsets to compute shifted centroid positions, cluster cells sharing nearby centroids, apply two-stage outlier filtering (cells must form valid clusters in both current and motion-predicted future frames), and smooth motion vectors within clusters using coefficient-of-variation thresholding to discard high-variance clusters.","core_discovery":"The paper's central discovery is that a learned centroid-offset prediction — an objectness prior indicating where each cell's parent object center lies — transfers from synthetic to real LiDAR data reliably enough to serve as a structural constraint on noisy motion pseudo-labels. When the teacher model generates motion predictions on real data, the objectness prior enables clustering cells into approximate rigid objects, filtering cells whose motion and centroid predictions are mutually inconsistent, and averaging motion vectors within each surviving cluster. This dual-path consistency check (motion branch and objectness branch must agree on cluster membership) suppresses the two dominant失败s","pith_inferences":["The large performance gap between Waymo and nuScenes may reflect differences in sensor configuration (32 vs 64 beams, different mounting, different density) affecting how well centroid-offset predictions learned on synthetic data transfer — a factor the paper does not isolate from the rigid-object assumption itself.","The objectness prior's success on static cells (where the method nearly matches oracle performance) versus its weaker performance on fast-moving cells suggests that the prior is most useful when motion is simple (zero or uniform) and degrades when motion is complex, which is precisely the regime where structural constraints would be most valuable for safety-critical applications.","If the centroid-offset prediction is the key transferable quantity, one could test this directly by training the objectness branch on synthetic data, evaluating centroid-offset accuracy on real data with available bounding-box labels (e.g., the Waymo detection annotations used to derive oracle motion labels), and correlating offset accuracy with motion prediction improvement — this would decouple ","The coefficient-of-variation threshold (T_CV = 10) for discarding high-variance clusters is a free parameter chosen on Waymo validation data; its sensitivity across datasets could indicate whether the method is learning a general principle or tuning to a specific distribution of object shapes and motion patterns."],"forward_implications":["If the objectness prior transfers reliably, autonomous driving systems could train motion predictors on entirely synthetic data with perfect labels, eliminating the costly detection-and-tracking pipeline currently needed to produce motion ground truth from real LiDAR.","The dual-path consistency principle (two independent prediction branches must agree before a pseudo-label is trusted) could generalize beyond motion prediction to any teacher-student domain adaptation setting where the teacher's labels are noisy — for instance, depth estimation, scene flow, or occupancy prediction under synthetic-to-real shifts.","The Motion4D dataset and its generation pipeline (static real backgrounds plus CAD assets on real trajectories, ray-traced with z-buffer occlusion) provide a template for generating labeled 4D LiDAR at scale, which could be extended to other perception tasks like detection or tracking.","The gap between Waymo results (closing 76.3% of the oracle gap on static cells) and nuScenes results (fast error 2.486 vs oracle 1.074, a 131% relative gap) suggests the rigid-object assumption's transferability is dataset-dependent, pointing to a need for deformable-object handling before the method generalizes universally."],"fun_headline_variants":["Learned objectness priors bridge synthetic-to-real gap in LiDAR motion prediction","Centroid-offset prediction transfers from synthetic to real LiDAR for motion filtering","Dual-path consistency check suppresses noisy motion pseudo-labels in real scenes","Objectness-aware clustering filters inconsistent motion cells across domain shift","Synthetic 4D LiDAR motion labels refined via objectness-guided cluster consistency"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The entire enhancement pipeline rests on the assumption that cells within a rigid object share consistent motions and a common centroid. If the centroid-offset predictions learned on synthetic data do not transfer to real data, or if real-world objects are deformable, partially observed, or have centroids that shift non-rigidly, the clustering, outlier filtering, and smoothing all degrade. The nuScenes results, where fast-group error remains 131% above the oracle, suggest the","fun_headline_variants_meta":{"raw":{"variants":["Learned objectness priors bridge synthetic-to-real gap in LiDAR motion prediction","Centroid-offset prediction transfers from synthetic to real LiDAR for motion filtering","Dual-path consistency check suppresses noisy motion pseudo-labels in real scenes","Objectness-aware clustering filters inconsistent motion cells across domain shift","Synthetic 4D LiDAR motion labels refined via objectness-guided cluster consistency","Motion and centroid branches must agree: a cross-domain filtering rule for LiDAR","Objectness priors from synthetic data constrain real-world motion pseudo-labels","Clustering cells by learned centroid offsets removes domain-shift noise in motion"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":802,"prompt_tokens":530,"completion_tokens":272,"prompt_tokens_details":null},"tokens_in":530,"tokens_out":272,"duration_ms":28196,"temperature":1.0,"reasoning_tokens":147,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T09:50:49.676867+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If centroid-offset predictions learned on synthetic Motion4D data produce systematically biased or high-variance offset vectors on real LiDAR (e.g., because synthetic object shapes, sizes, or LiDAR reflection patterns differ from real ones), the clustering step would group cells incorrectly, the outlier filter would discard valid cells or retain noisy ones, and the smoothing would average inconsistent motion vectors — producing pseudo-labels worse than the raw teacher predictions. The nuScenes fast-group results (2.486 vs oracle 1.074) are consistent with this failure mode being partially活性","supporting_citations":[],"review_version":1}