{"id":"ebb61910-afd8-4206-ad22-269999746e51","arxiv_id":"2506.09417","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ODG uses separate static and dynamic Gaussian query sets, refined coarse-to-fine, plus rendering supervision, and reports state-of-the-art occupancy prediction on Occ3D-nuScenes and Occ3D-Waymo.","lead":"This paper introduces ODG, a 3D occupancy prediction system that represents driving scenes with two separate sets of Gaussians, one for static and one for dynamic objects. It reports state-of-the-art accuracy on the Occ3D nuScenes and Waymo benchmarks, with a faster small model that trails the large model in accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA comparison is confounded by training schedule: at 24 epochs the complete ODG-T variant reaches only 32.82 mIoU (Tab. 4), below OPUS-T's 33.2, while the reported 38.18/35.54 numbers use a 100-epoch schedule.","rationale":"The reader's weakest assumption concerns Gaussian scale/rotation/opacity being optimized only through the rendering loss. That is a real limitation and is explicitly acknowledged in the paper, but it is not the most load-bearing vulnerability. The central empirical claim is that ODG sets new state-of-the-art results against OPUS. For that claim to hold, the comparison must be apples-to-apples. The paper's own numbers strongly suggest it is not: the complete 24-epoch ablation variant (32.82 mIoU) falls below the OPUS-T baseline (33.2 mIoU) it claims to beat, and the only reported difference that can explain the improvement to 35.54 is the 100-epoch schedule. Because training budget is a first-order factor in sparse transformer occupancy models, the +1.98 and +2.34 mIoU margins cannot currently be attributed to the proposed dual-Gaussian or hierarchical designs. The box supervision and rendering supervision further confound attribution. This is not an accusation of misconduct; it is a request for a controlled comparison. Given that the paper's headline claim is 'new state-of-the-art results' and that claim is not supported under matched conditions, the current manuscript's central empirical statement should be rejected as stated, with the possibility of revision after a fair retraining comparison.","tokens_in":14113,"tokens_out":13036,"duration_ms":143518,"concrete_test":"Run a matched training-budget comparison on Occ3D-nuScenes: train ODG-T and OPUS-T (and ideally ODG-L and OPUS-L) at both 24 and 100 epochs with the same batch size, image resolution, and data sampling, reporting mIoU and RayIoU. If OPUS-T at 24 epochs meets or beats ODG-T at 24 epochs while ODG-T wins only at 100 epochs, the headline SOTA claim is a training-schedule artifact. Also ablate the box loss in Eq. 17 to quantify how much of any remaining gain comes from the extra 3D box supervision.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing issue is that the headline SOTA comparison is not matched in training budget. Section 4.1 says ODG is trained with batch size 8 for 100 epochs, but the paper never discloses the training schedule of the OPUS entries in Tabs. 1 and 3. The paper's own ablations in Sec. 4.4 are run for 24 epochs, and the complete ODG-T variant in Tab. 4 reaches only 32.82 mIoU, below OPUS-T's 33.2 mIoU in Tab. 1. At 100 epochs ODG-T reaches 35.54, a +2.72 gain from training alone. Hence a substantial part of the claimed +2.34 (ODG-T vs OPUS-T) and +1.98 (ODG-L vs OPUS-L) margins may be an artifact of a roughly 4x-longer schedule rather than of the dual-Gaussian/hierarchical design. Additionally, Eq. 17 adds 3D box supervision and Eq. 19 adds LiDAR-projected depth/semantic supervision, neither of which OPUS uses; the box loss is not ablated, so its contribution is unknown. The Limitations section honestly discloses only the Gaussian-parameter issue, not this comparison confound.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ODG, a camera-based 3D semantic occupancy prediction method that represents the scene with two sets of sparse Gaussian queries, one for static background and one for dynamic agents. Dynamic queries additionally carry 3D box attributes and a velocity vector, and the two query sets interact through a self-attention mechanism over their concatenated features. Gaussians are predicted in a hierarchical coarse-to-fine transformer, with supervision from Chamfer distance and focal loss on occupied voxel centers and class scores, box losses from dynamic queries, and a rendering loss that uses 3D Gaussian splatting to render depth and semantic maps from the predicted Gaussians. Experiments on Occ3D-nuScenes and Occ3D-Waymo report state-of-the-art mIoU and RayIoU numbers, and ablations study motion compensation, query attention, and rendering supervision. The paper is clearly written and the design is well motivated.","tokens_in":14422,"tokens_out":5969,"duration_ms":59124,"significance":"The dual static/dynamic decomposition is a sensible inductive bias for driving scenes, and the coarse-to-fine Gaussian growth is a plausible extension of sparse-query occupancy models. The component-wise ablations on public benchmarks give initial support for each design choice, and the paper honestly discloses in the Limitations section that scale, rotation, and opacity are optimized only through the rendering loss. If the headline gains survive a controlled training-budget comparison and an ablation of the extra supervision signals, ODG would be a meaningful advance for efficient camera-based occupancy prediction. However, the current manuscript does not yet establish that the proposed architecture, rather than longer training and additional 3D box and LiDAR supervision, is responsible for the reported state-of-the-art margins.","major_comments":[{"comment":"The headline SOTA comparisons are not matched in training budget. ODG is trained with a global batch size of 8 for 100 epochs, while the ablations in Sec. 4.4, including the complete ODG-T row in Tab. 4, use 24 epochs. At 24 epochs ODG-T reaches 32.82 mIoU, which is below OPUS-T's 33.2 mIoU in Tab. 1. The training schedules for the OPUS entries in Tabs. 1 and 3 are not disclosed, so the relative contributions of architecture versus longer training to the claimed +2.34 and +1.98 mIoU margins cannot be assessed. Please match schedules, report OPUS numbers under the identical 100-epoch protocol, or provide an ablation that shows the 24-epoch ODG/OPUS comparison; also disclose the OPUS training schedules used in the tables.","section":"Sec. 4.1, Tables 1/3/4"},{"comment":"The method adds two supervision signals, a 3D box loss and a rendering loss using LiDAR-projected depth and semantic labels, that are not used by OPUS. While the rendering supervision itself is ablated in Tab. 4 (+0.69 mIoU), the box loss Lbox is never ablated, so its contribution to the reported gains is unknown. To establish that the dual-Gaussian/hierarchical design, rather than the additional annotations, drives the improvement, please include an ablation without Lbox and, if possible, an ablation that uses the same set of labels as OPUS.","section":"Sec. 3.5, Eqs. (17) and (19)"},{"comment":"The efficiency claim is not supported by the reported FPS numbers. ODG-L runs at 4.9 FPS while OPUS-L runs at 7.2 FPS, and ODG-T runs at 20.1 FPS versus 22.4 FPS for OPUS-T; both ODG variants are slower. Reporting parameter counts, FLOPs, and memory usage would give a more informative comparison, since FPS alone is hardware- and implementation-dependent. Please either revise the 'low inference cost' claim or provide a more complete efficiency evaluation.","section":"Sec. 4.1, Table 1"},{"comment":"The coarse-to-fine growth schedule K_l is never specified. Eq. (4) defines K_l as the number of Gaussians per query at stage ℓ, and Eqs. (9)-(10) assert K_{ℓ-1} < K_ℓ, but the actual values (e.g., K_1 through K_6) and the mechanism that expands the Gaussians between layers are absent from the implementation details. Without this schedule, the hierarchical capacity of the model cannot be reproduced or compared with prior single-stage methods. Please provide the K_l values for all stages and the rule that maps the transformer layers to the increasing number of Gaussians.","section":"Sec. 3.2 and Sec. 4.1 (Implementation Details)"},{"comment":"All ablations and the final model are single runs with no variance information. Several differences are small relative to typical seed-to-seed variation (e.g., the mIoU difference between Cross Attn and Self Concat Attn in Tab. 5a is 0.18, and the RayIoU difference in Tab. 5b is 0.5). Adding standard deviations over at least three seeds would make the component-level claims more convincing.","section":"Sec. 4.4"}],"minor_comments":[{"comment":"In the definition of Gs and Gd, the subscript K_l is used, but K_l is stage-dependent; please clarify that the notation applies per stage ℓ.","section":"Eq. (4)"},{"comment":"The layer index ℓ is used both for transformer layers and for stages; clarify the relationship between the L transformer layers and the K_l growth schedule, e.g., whether each layer corresponds to exactly one stage.","section":"Eqs. (9)-(10)"},{"comment":"The loss includes a Chamfer distance term for the initial Gaussian means G:µ,0, but Sec. 3.2 states that Gaussian means are initialized from U[0,1]; please explain how the initial means are obtained or decoded before the first layer.","section":"Sec. 3.5, Eq. (16)"},{"comment":"The abbreviations 'Ego Comp.' and 'Dyn. Comp.' are not defined in the table caption; please define them as ego-motion compensation and dynamic-object motion compensation.","section":"Tab. 5b"},{"comment":"The implementation details list the values of S, D, and L but not K_l; please include the exact K_l values to make the coarse-to-fine design reproducible.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a preprint under review and does not yet provide code. The missing K_l schedule and the undisclosed OPUS training budgets make the current results difficult to verify independently. I recommend asking the authors for matched-budget comparisons, an ablation of the box loss, and seed variance estimates before a final decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the architecture is a real addition to the sparse-Gaussian occupancy line, but the headline SOTA comparison is not matched on training budget, and the paper's own ablations imply it loses to OPUS at equal epochs. That should be the first thing the authors fix.\n\nWhat is genuinely new: the dual static/dynamic Gaussian query split, the hierarchical coarse-to-fine Gaussian refinement across transformer layers, and the multi-stage rendering supervision. The ablations in Tab. 4 are clean: motion compensation helps, query attention helps, rendering supervision helps, and each component is honestly isolated. The Limitations section also flags, correctly, that scale/rotation/opacity are only optimized via rendering loss.\n\nWhere it gets soft. The paper trains all models for 100 epochs but never discloses OPUS's schedule. If OPUS uses the standard 24-epoch training, the comparison is unfair: the complete ODG-T at 24 epochs is 32.82 mIoU, below OPUS-T's 33.2 mIoU. That means the +2.34 margin at 100 epochs could be mostly or entirely the longer schedule, not the dual-Gaussian design. The stress-test note on this holds up. In addition, the 3D box loss in Eq. 17 is never ablated, so its contribution is unknown; the FPS claim applies to ODG-T, not the SOTA ODG-L, which runs at 4.9 FPS versus OPUS-L's 7.2; and there are no error bars or significance tests. Minor issues: the K_l coarse-to-fine schedule is never specified, and some class-wise numbers on Waymo (e.g., the pedestrian jump) deserve a second look.\n\nBottom line: the core idea is worth refereeing, and the architecture is plausible, but the empirical claim as stated is not supported by the evidence in the draft. The fix is straightforward: match training schedules, ablate the box loss, and report standard deviations. If the margin survives matched-schedule comparison, this is a solid paper. As it stands, I would not cite it as SOTA until that is done.","headline":"The dual static/dynamic Gaussian architecture is a genuine contribution, but the headline SOTA claim is confounded by an unmatched 100-epoch training schedule versus OPUS.","tokens_in":14950,"tokens_out":3510,"would_cite":false,"duration_ms":35238,"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":"ODG represents a driving scene as two separate sets of 3D Gaussians, one static and one dynamic, and reports new state-of-the-art occupancy prediction results on both Occ3D benchmarks.","keywords":["3D occupancy prediction","Gaussian splatting","sparse queries","static-dynamic decomposition","coarse-to-fine refinement","autonomous driving","camera-only perception","Occ3D benchmark"],"falsifier":"Re-run ODG on Occ3D-nuScenes with the rendering loss term $L_r$ disabled and compare against the full model; if the mIoU drop is not close to the reported +0.69 and RayIoU drop is not close to +0.70, the paper's attribution of accuracy to rendering supervision is falsified. A second check is to replace LiDAR-projected labels with monocular predicted depth and semantics: if performance holds, the dependence on LiDAR-derived labels is not essential.","tokens_in":13941,"feed_emoji":"🚗","tokens_out":8774,"duration_ms":76333,"temperature":0.7,"pith_summary":"ODG claims that camera-only 3D occupancy prediction improves when the scene is represented not by one set of sparse queries but by two: one set of 3D Gaussians dedicated to static background and another, augmented with bounding-box and velocity attributes, dedicated to dynamic agents. The paper argues that a single query set spreads capacity too thinly across objects that behave very differently, and that a coarse-to-fine hierarchy of transformer layers lets the model afford far more Gaussians than prior single-transformer designs. It also uses 3D Gaussian splatting to render depth and semantic maps at every stage, supervising those renders with LiDAR-projected labels to inject pixel-level alignment. On Occ3D-nuScenes and Occ3D-Waymo, ODG reports new state-of-the-art mIoU of 38.18 and 21.35, surpassing the previous best by 1.98 and 2.35 points respectively while keeping inference cost competitive. The core claim a sympathetic reader should take away is that static/dynamic decomposition plus multi-stage rendering supervision is the right way to scale sparse occupancy prediction.","feed_headline":"Two Gaussian query sets beat prior 3D occupancy predictors","feed_subtitle":"Static and dynamic Gaussians plus rendering supervision hit 38.18 mIoU on Occ3D-nuScenes, topping OPUS-L by 1.98.","key_machinery":"The load-bearing mechanism is the dual Gaussian query representation: two sets of 3D Gaussians, where static queries carry $\\{\\mu,s,r,\\sigma\\}$ and dynamic queries append the box-velocity attribute $b=[l,w,h,\\theta,v_x,v_y,v_z]$. A hierarchical Gaussian transformer $T_\\ell$ refines Gaussian means through coarse-to-fine stages, DaS attention exchanges information between static and dynamic queries, and 3D Gaussian splatting renders depth and semantic maps per camera for multi-stage supervision. These pieces work together to let a sparse set predictor afford a much larger Gaussian count and to make the non-mean Gaussian properties learn from pixel-level 2D labels rather than only from 3D occupancy.","core_discovery":"On the paper's own terms, the central discovery is that driving scenes admit a universal static/dynamic decomposition, and encoding it directly into the query structure pays off. ODG initializes two sets of Gaussian queries, static $\\{g^s_{i,k}\\}$ and dynamic $\\{g^d_{j,k}\\}$, where dynamic queries append the box-velocity attribute $b=[l,w,h,\\theta,v_x,v_y,v_z]$ to the standard Gaussian parameters $\\{\\mu,s,r,\\sigma\\}$. A Dynamic-and-Static (DaS) attention concatenates both query families before self-attention, and a hierarchy of transformer layers refines Gaussian means coarse-to-fine, with $K_{\\ell-1}