{"id":"2f3ec0a1-f777-4f11-8434-f29827d0d6be","arxiv_id":"2506.14130","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"KDMOS distills knowledge from the MambaMOS teacher into a lightweight BEV student, using class-decoupled and label-weighted distillation, reaching 78.8% IoU on SemanticKITTI-MOS.","lead":"A team proposes a knowledge distillation method that lets a heavy, accurate LiDAR motion segmentation model teach a lightweight bird's-eye-view model, improving accuracy while running at 40 FPS. The approach could make real-time moving object segmentation more practical for autonomous driving.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WDCD's input alignment is unspecified: Eq. (8) is defined on point logits while the student emits a BEV grid, and the four-class probabilities required by Eq. (5) are never sourced from the binary MOS teacher.","rationale":"The paper's contribution is a distillation loss, so the loss must be well-defined on the actual tensors. The reader's weakest assumption already identified the point-to-BEV ambiguity; I agree and add that the four-class logit assumption in Eq. (5) is equally unexplained. Both are load-bearing: if the alignment is ad hoc or the teacher has only binary logits, the reported gains could come from the student's own loss or from a different distillation signal than the one described. This is not a disagreement with the community's consensus; it is an internal discrepancy between Sec. III-A's BEV representation and Sec. III-C's per-point, four-class formulation. The Eq. (8) algebra itself is correct as a rearrangement, and the ablation table is internally consistent, so the concern is about missing linkage, not fabrication. The GitHub repository may settle this, which is why a code-level check is the right test. Since the paper does not currently allow a reader to reconstruct the central loss, the conditional verdict remains appropriate; I would not accept without this clarification.","tokens_in":9780,"tokens_out":7698,"duration_ms":83363,"concrete_test":"Inspect the released KDMOS code's loss function and reproduce the WDCD computation on one SemanticKITTI validation batch. Verify (a) that teacher logits have four classes, and (b) the exact indexing that aligns teacher point logits to student BEV cells, including the handling of cells containing multiple points or scans with conflicting labels. Then rerun the ablation with the mapping deliberately changed, for example nearest-cell versus max-occupancy aggregation; if validation IoU moves by more than 1 point, or if four-class teacher logits cannot be constructed, the missing specification is load-bearing and the central claim should be considered unverified until corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the student's 78.8% test IoU follows from WDCD depends on the loss in Sec. III-C being computable. As written, Eq. (8) is a per-point KL over four logits, but the student is a BEV model: Sec. III-A describes 2D grids, not point predictions. The text never states whether teacher point logits are rendered into BEV cells, whether student BEV logits are projected back to points, or what happens when multiple points or multiple scans map to one cell. The same gap applies to the class space: Eq. (5) requires four logits (unlabeled, static, movable, moving) for both teacher and student, yet the teacher is a frozen MambaMOS model pretrained for binary MOS, with no description of how its output supplies these four logits. If the teacher output is binary, the non-target KL term in Eq. (8) is degenerate and WDCD reduces to an ordinary binary KL, so the reported ablation gains (+1.3 to +2.9 IoU) cannot be attributed to the described mechanism. This is a reproducibility and correctness risk, not a style issue. The GitHub release could resolve it, but the paper alone does not.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KDMOS, a knowledge-distillation framework for LiDAR moving-object segmentation. A frozen MambaMOS teacher distills logits into a BEV-based MotionBEV student via a weighted decoupled class distillation (WDCD) loss, together with a DySample-based upsampling module. The authors report 79.4% validation and 78.8% test IoU on SemanticKITTI-MOS, 40 FPS inference, and a 7.69% parameter reduction over MotionBEV, plus generalization results on Apollo. The algebraic derivation in Eqs. (7)-(8) is correct, and the ablations in Tables III-V are internally consistent, but several implementation details required to compute the distillation loss are missing.","tokens_in":9987,"tokens_out":5582,"duration_ms":58655,"significance":"If the reported numbers are reproducible, the contribution is practically significant: it would be the first logits-based KD for MOS, achieving near-MambaMOS accuracy at real-time speed with a simple, parameter-free-at-inference loss. The derivation in Eqs. (7)-(8) is correct and the ablation tables support the modular claims. However, the current manuscript does not specify how teacher point logits are aligned with student BEV outputs or how a binary teacher provides four-class logits, so the central mechanism cannot be verified from the text alone.","major_comments":[{"comment":"The distillation loss is written as a per-point KL between teacher probability p_T and student p_S, but the teacher operates on unordered 4D point clouds while the student outputs a BEV grid. The manuscript never states how teacher point logits are associated with student BEV cells, how student BEV logits are converted back to per-point predictions, or how collisions (multiple points per cell, multiple scans per cell) are resolved. Without this alignment, Eq. (8) is not computable from the provided text, so the ablation gains in Table III cannot be independently reproduced or verified.","section":"Section III-C, Eq. (8)"},{"comment":"Eq. (5) defines four class probabilities (unlabeled, static, movable, moving), but the teacher is described in Section III-B as MambaMOS, a binary moving-object segmentation model for SemanticKITTI-MOS. The paper does not specify how the frozen teacher supplies four logits per point. If the teacher provides only binary moving/non-moving logits, then the non-target distribution \\hat p has only one effective class and the NCKD term in Eq. (8) is degenerate; the claimed benefit of non-target class distillation would then not be attributable to the stated mechanism. Please state the teacher output space and, if necessary, the mapping from binary scores to the four classes.","section":"Section III-C, Eq. (5)"},{"comment":"All quantitative results are single training runs; no standard deviations or number of seeds are reported. The reported gains are small (e.g., +1.3 IoU in Table III, +2.9 in validation), so without variance estimates or repeated runs it is unclear whether the differences are significant. Additionally, the hyperparameters alpha=0.25, gamma=0.25, beta, and window sizes N, Q1, Q2 are fixed without sensitivity analysis; a figure or table showing IoU as a function of these parameters would strengthen the claim that the chosen settings are not cherry-picked.","section":"Section IV-A, Tables IV and V"},{"comment":"The definition of W_i is ambiguous: 'Content[label]' is not given a mathematical expression, and WDCD = DCD / W_i implies inverse-frequency weighting, whereas the text describes 'label-assigned weighting' without specifying the direction. Because this term is central to the proposed loss, please define Content[label] explicitly and state whether division (rather than multiplication) by this ratio is intended; otherwise the implementation cannot be reproduced.","section":"Section III-C, Eq. (10)"}],"minor_comments":[{"comment":"The table title 'CROSSVALIOU' should read 'Cross-validation performance' or similar; please fix the typo.","section":"Table II"},{"comment":"The item numbered '1)' at the start of Section III-B is a formatting fragment; please move it into a proper enumerated list.","section":"Section III-B"},{"comment":"The last row 'KDMOS40 254.08 15' should be formatted as 'KDMOS 40 25 4.08 15' so that the FPS, ms, params, and size columns are legible.","section":"Table VI"},{"comment":"The relationship between 'Content[label]' and the per-frame class frequencies should be stated with an explicit formula, since the current text only says 'ratio of points from different categories in the ith frame'.","section":"Section III-C, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The missing teacher-student alignment and four-class logit mapping are the main risks. If the authors can point to exact code lines or add a precise algorithm box, the issues are local. If the supplied code does not implement Eq. (8) as written, the paper's central claim may not survive. I would require the code and a written description as part of the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a decent, workmanlike KD paper for LiDAR MOS. The new thing is specific: decoupling moving vs non-moving classes in a logit-based distillation loss and weighting by label frequencies, applied to teach a BEV student from a MambaMOS teacher. The DKD derivation in Eqs. 7-8 is correct, the ablations are consistent, and the headline numbers (78.8% test IoU, 40 FPS, 7.69% fewer params than MotionBEV) are genuinely useful if they reproduce. The paper also deserves credit for shipping code and for testing the module on other MOS baselines.\n\nThat said, the load-bearing part of the method is underspecified. Eq. 8 is per-point KL over four logits, but the student outputs a BEV grid; the paper never says how teacher point logits are aligned to student cells, what happens when multiple points or scans fall in one cell, or whether student BEV logits are projected back to points. More seriously, Eq. 5 requires four class probabilities from both teacher and student, but MambaMOS is a binary MOS model. There is no description of how four logits are obtained from it. If the teacher only gives two outputs, the non-target term in Eq. 8 is degenerate and WDCD collapses to ordinary binary KL, meaning the +1.3 to +2.9 IoU gains attributed to the decoupling cannot be traced to the described mechanism. The GitHub release could resolve this, but the paper alone does not.\n\nOther, smaller issues: all results are single runs without variance, the hyperparameters (alpha, gamma, beta) are fixed with no sensitivity analysis, and the fact that the teacher and several cited baselines (MF-MOS, CV-MOS, MambaMOS) come from the same lab is not disclosed. None of these are fatal, but they cap confidence. The Apollo results are also presented without the same scrutiny as the main benchmark.\n\nWho is this for? Researchers working on real-time MOS deployment or KD for LiDAR perception. A serious referee should get it because the result is useful and the idea is plausible, but the reviewer should ask for a precise description of the logit-space alignment and the teacher's output space before the method can be reproduced from text.\n\nRecommendation: send to peer review, with a request for major revision clarifying the distillation geometry and class space.","headline":"Useful incremental KD result for MOS, but the distillation loss's geometric and class-space alignment is underspecified enough that the core mechanism may not be what it claims.","tokens_in":10605,"tokens_out":2104,"would_cite":false,"duration_ms":21328,"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 distilled BEV student reaches 78.8% IoU on SemanticKITTI-MOS at 40 FPS, closing most of the gap to a slow point-cloud teacher.","keywords":["moving object segmentation","knowledge distillation","LiDAR point cloud","bird's-eye view","class imbalance","decoupled distillation","real-time inference","SemanticKITTI-MOS"],"falsifier":"Compute WDCD on the validation set with the teacher logits attached to randomly chosen grid cells rather than the actual ones; if the student's validation IoU stays near 79.4%, then the alignment between teacher points and student cells is not doing the work, and the distillation signal could be arbitrary.","tokens_in":9510,"feed_emoji":"🚗","tokens_out":6963,"duration_ms":68530,"temperature":0.7,"pith_summary":"The paper tries to prove that motion segmentation for autonomous driving does not have to trade accuracy for speed. It trains a fast bird's-eye-view student by distilling the outputs of a powerful but slow non-projection point-cloud teacher, with a loss shaped for the extreme imbalance between moving and static points. In the paper's experiments, the student reaches 78.8% test IoU on SemanticKITTI-MOS at 40 FPS, just 1.3 points below the heavy teacher, while also reducing parameters relative to its BEV baseline. If true, this gives a practical recipe for deploying motion segmentation in real time without sacrificing most of the accuracy of the most accurate offline models.","feed_headline":"Distilled BEV model hits 78.8% IoU at 40 FPS","feed_subtitle":"A small student closes most of the gap to a costly point-cloud teacher on motion segmentation.","key_machinery":"The central mechanism is Weighted Decoupled Class Distillation (WDCD), a logits-only loss that rewrites ordinary KL distillation as a binary target/non-target term plus a non-target distribution term, then further branches by semantic class. For moving points, both terms are distilled; for the abundant non-moving class, only the non-target term is kept, and the loss is divided by class-content weights so the imbalance does not drown out rare moving points. The second piece of machinery is DySample, a point-sampling dynamic upsampler that replaces fixed convolutional upsampling, shrinking the student to 4.08M parameters.","core_discovery":"KDMOS claims that a fast BEV-projection student can inherit most of the accuracy of a slow 4D point-cloud teacher through logits-only distillation, that is, by training on the teacher's final per-class scores. On SemanticKITTI-MOS, this student scores 79.4% validation IoU and 78.8% test IoU, which is 3.9 and 2.9 points above the MotionBEV baseline and only 2.9 and 1.3 points below the frozen MambaMOS teacher, while running at 40 FPS and shrinking parameters by 7.69%. The paper attributes the gain to a class-aware distillation loss that avoids letting the abundant non-moving class dominate the learning signal.","pith_inferences":["The moving-versus-non-moving decoupling should transfer to any segmentation task with extreme class imbalance, since the dominating easy class can be given only the non-target distillation term.","A cheap way to stress-test KDMOS is to inspect the teacher-to-BEV projection: counting how many teacher points share one student cell, and whether their labels agree, would tell whether the point-wise KL is computed on clean correspondences.","Because only logits are transferred, the teacher's learned temporal coupling enters the student only through final probabilities; a feature-level motion distillation could close the remaining 1.3-point gap, at the cost of architectural alignment.","The method's value in deployment is exactly the student's 40 FPS, since the teacher is needed only at training time; this is a direct consequence of choosing a logits-based scheme."],"forward_implications":["A BEV student trained with WDCD reaches 79.4% validation and 78.8% test IoU on SemanticKITTI-MOS, improving over MotionBEV by 2.9 to 3.9 points while staying real-time.","The student runs at 40 FPS with 4.08M parameters, a 7.69% parameter cut and faster inference than the 34 FPS MotionBEV baseline.","WDCD is logits-only, so it adds no parameters, and the paper shows it lifts other MOS baselines when appended to them.","Dropping the target-class term for non-moving points avoids a harmful distillation effect that otherwise costs 0.3% IoU, so the decoupling is a correctness move rather than just a speed trick.","On Apollo, the method reaches 68.2% validation IoU without domain adaptation, supporting the claim that the distilled student generalizes across environments."],"supporting_citations":[{"why":"The BEV baseline and input representation KDMOS starts from; supplies MotionBEV's validation/test IoU and parameter count that WDCD must beat.","marker":"[8]"},{"why":"The frozen non-projection teacher whose per-point logits are distilled; supplies the 4D point-cloud spatio-temporal representation and the MambaMOS teacher.","marker":"[11]"},{"why":"Introduced logits-based knowledge distillation, the method family WDCD belongs to.","marker":"[12]"},{"why":"Decoupled knowledge distillation gives the target/non-target KL decomposition that WDCD re-decouples by moving versus non-moving classes.","marker":"[16]"},{"why":"Provides the DySample dynamic upsampling module that reduces the student's parameter count.","marker":"[21]"},{"why":"Defines the SemanticKITTI dataset from which the MOS benchmark and evaluation protocol are drawn.","marker":"[25]"},{"why":"Introduces the SemanticKITTI-MOS moving-object segmentation benchmark and the training/validation setup adopted here.","marker":"[6]"},{"why":"Supplies the Apollo dataset used to test cross-environment generalization without domain adaptation.","marker":"[17]"}],"fun_headline_variants":["Logits-only distillation gives BEV motion segmentation 40 FPS speed","Distilled BEV student hits 78.8% IoU with 7.69% fewer parameters","Class-aware KD lets fast BEV model match point-cloud teacher","Motion segmentation distillation: student BEV beats MotionBEV by 3.9 IoU","Decoupled motion classes make distillation boost BEV MOS accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the teacher's per-point predictions can be meaningfully paired with the student's BEV cells for a per-point loss, yet the paper does not explain which point goes to which cell when multiple scans land in one cell.","fun_headline_variants_meta":{"raw":{"variants":["Logits-only distillation gives BEV motion segmentation 40 FPS speed","Distilled BEV student hits 78.8% IoU with 7.69% fewer parameters","Class-aware KD lets fast BEV model match point-cloud teacher","Motion segmentation distillation: student BEV beats MotionBEV by 3.9 IoU","Decoupled motion classes make distillation boost BEV MOS accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000871,"raw_usage":{"total_tokens":3750,"prompt_tokens":901,"completion_tokens":2849,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2747}},"tokens_in":517,"tokens_out":2849,"duration_ms":19675,"temperature":1.0,"reasoning_tokens":2747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:17:45.885884+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute WDCD on the validation set with the teacher logits attached to randomly chosen grid cells rather than the actual ones; if the student's validation IoU stays near 79.4%, then the alignment between teacher points and student cells is not doing the work, and the distillation signal could be arbitrary.","supporting_citations":[{"cited_title":"Motionbev: Attention- aware online lidar moving object segmentation with bird’s eye view based appearance and motion features,","cited_arxiv_id":null,"evidence_quote":"The BEV baseline and input representation KDMOS starts from; supplies MotionBEV's validation/test IoU and parameter count that WDCD must beat."},{"cited_title":"Mambamos: Lidar-based 3d moving object segmentation with motion-aware state space model,","cited_arxiv_id":null,"evidence_quote":"The frozen non-projection teacher whose per-point logits are distilled; supplies the 4D point-cloud spatio-temporal representation and the MambaMOS teacher."},{"cited_title":"Learning to upsample by learning to sample,","cited_arxiv_id":null,"evidence_quote":"Provides the DySample dynamic upsampling module that reduces the student's parameter count."},{"cited_title":"Semantickitti: A dataset for semantic scene un- derstanding of lidar sequences,","cited_arxiv_id":null,"evidence_quote":"Defines the SemanticKITTI dataset from which the MOS benchmark and evaluation protocol are drawn."},{"cited_title":"Moving object segmentation in 3d lidar data: A learning-based approach exploiting sequential data,","cited_arxiv_id":null,"evidence_quote":"Introduces the SemanticKITTI-MOS moving-object segmentation benchmark and the training/validation setup adopted here."},{"cited_title":"L3-net: Towards learning based lidar localization for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Supplies the Apollo dataset used to test cross-environment generalization without domain adaptation."}],"review_version":1}