{"id":"b6ca936c-750a-4492-970b-93be282f984c","arxiv_id":"2509.03999","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SliceSemOcc improves 3D semantic occupancy prediction by slicing voxel features into global and local height bands and applying per-height channel attention, yielding modest mIoU gains on nuScenes benchmarks.","lead":"This paper introduces SliceSemOcc, a 3D scene understanding method that slices voxel features along the height axis and applies per-height attention to better recognize small objects like pedestrians and traffic cones. Trained on camera and LiDAR data, it reports improved semantic occupancy prediction accuracy on two nuScenes benchmarks, though gains over the strongest prior methods are small.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline gain over M-CONet is confounded by an uncontrolled baseline: Table 3's no-slice baseline already exceeds M-CONet by 2.3 mIoU, so the claimed 3.5-point gain is not solely due to vertical slicing.","rationale":"The reader's verdict (CONDITIONAL) is appropriate, but the most load-bearing concern is not primarily the hand-selected slice boundaries; it is the uncontrolled baseline used for the headline comparison. The internal ablation provides a control that already outperforms the published M-CONet by 2.3 points, meaning the 14.2% relative gain and the large per-class gains in Section 4.2 conflate the proposed modules with unrelated implementation changes. This directly threatens the paper's central claim as stated. However, the method is coherent, the modules are plausible, and the internal ablation does show a consistent, if smaller, improvement (1.23 mIoU). The paper should be accepted only if the authors release code and provide a controlled M-CONet baseline under their own pipeline, or clearly restate gains relative to their internal baseline. The reader identified a different but related weakness (validation-informed slice boundaries); I partially agree, since both concerns affect the reliability of the reported improvement, but the baseline mismatch is the more consequential one. No verdict change is needed beyond the existing conditional recommendation, because the paper can be made correct with additional controlled experiments and disclosure.","tokens_in":11926,"tokens_out":8159,"duration_ms":78671,"concrete_test":"Train the exact same pipeline as SliceSemOcc on nuScenes-SurroundOcc with the VSF and SEAttention3D modules removed (the Table 3 row-1 configuration), using identical losses, view transformer, backbone, and training schedule. If this control yields ~26.98 mIoU, then the true contribution of the proposed modules is ~1.2 mIoU and the headline comparison to M-CONet must be recomputed from this controlled baseline; if it reproduces M-CONet's 24.7 mIoU, then Table 3's baseline discrepancy needs a concrete explanation. This test would also let one verify whether the local slice boundaries were tuned on the validation split by abalating boundaries chosen from the training split only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central comparison (Table 1) reports SliceSemOcc at 28.2 mIoU versus M-CONet at 24.7 on nuScenes-SurroundOcc, a 3.5-point gain. But Table 3 shows that the authors' own baseline with both global and local slicing disabled reaches 26.98 mIoU. That is 2.28 points above the published M-CONet number. Section 4.1 lists implementation choices beyond the proposed modules—InverseMatrixVT3D projection, FPN, VoxelNet, and focal+Lovasz+affinity losses—so the 2.28-point gap likely reflects differences in training recipe and view transformation, not vertical slicing. The actual contribution of the proposed modules, measured internally, is 26.98 -> 28.21 (+1.23 mIoU, ~4.6% relative), not the claimed 14.2% relative improvement. The small-object percentage gains in Section 4.2 are computed relative to M-CONet's published class-wise numbers and inherit the same confound. The paper never presents a directly comparable M-CONet baseline trained under its own pipeline. This makes the headline claim that vertical slicing significantly improves over M-CONet unsupported as stated; what is supported is a modest improvement over an unreported, stronger internal baseline. The reader's concern about hand-selected slice boundaries is real, but the baseline mismatch is more load-bearing because it affects the headline gain itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SliceSemOcc, a multimodal (camera + LiDAR) 3D semantic occupancy prediction framework. The core ideas are (i) a Vertical Slice Fusion module that extracts features from global and local height slices of the voxel volume, (ii) a per-height channel-attention module (SEAttention3D) that preserves height resolution during pooling, and (iii) a global–local cross-attention fusion. Experiments on nuScenes-SurroundOcc and nuScenes-OpenOccupancy report mIoU of 28.2 and 22.9, respectively, with ablations claiming that each component contributes positively. The paper also reports a memory overhead of 0.6 GB for the full module.","tokens_in":12240,"tokens_out":3937,"duration_ms":36888,"significance":"If the reported internal improvements are robust, the work provides a useful, incremental architectural contribution: height-aware channel attention and dual-scale vertical slicing are intuitive and align with observed height distributions of small objects. The component-wise ablations (Tables 3–4, Section 4.3) are a strength, as is the explicit reporting of memory cost and the candid discussion of degraded flat-ground categories. However, the headline comparison is confounded by an uncontrolled baseline, the slice boundaries are derived from the same dataset on which the method is evaluated, and the second dataset result ties an existing method (OccLoff) without acknowledgement. No code, training hyperparameters, or multiple-seed statistics are provided, making the small reported gains difficult to verify.","major_comments":[{"comment":"The central claim of a 14.2% relative mIoU gain over M-CONet (24.7 → 28.2) is confounded by an uncontrolled baseline. Table 3 shows that the authors' own no-slice baseline achieves 26.98 mIoU, which is 2.28 points above the published M-CONet number. Section 4.1 lists many implementation differences from M-CONet (InverseMatrixVT3D projection, FPN, VoxelNet, focal+Lovasz+affinity losses), so the gap between 26.98 and 24.7 likely reflects training recipe and view-transformation differences, not vertical slicing. The internal contribution of the proposed modules is therefore 26.98 → 28.21, i.e., +1.23 mIoU (~4.6% relative), not 14.2%. The small-object gains in Section 4.2 are also computed relative to M-CONet's published class-wise numbers and inherit the same confound. The paper must either train M-CONet under its own pipeline or present the internal ablation as the primary result.","section":"§4.2, Tables 1 and 3"},{"comment":"The local slice boundaries [-5,-3], [-3,-2], [-2,-1], [-1,0], [0,1], [1,3] m are explicitly chosen based on the object height distribution of nuScenes-SurroundOcc (Fig. 1) and then evaluated on the validation set of the same dataset. The paper never demonstrates that these boundaries were fixed using only the training split, nor does it test transfer to another dataset. Table 4 shows that the boundary choice matters (28.21 vs. 27.67/27.75), so this is not a negligible design detail. The authors should state whether the boundaries were computed from training statistics only, and provide sensitivity/transfer experiments (e.g., applying the same boundaries to nuScenes-OpenOccupancy or reporting dataset-specific boundaries).","section":"§3.4, Fig. 1, Table 4"},{"comment":"On nuScenes-OpenOccupancy, SliceSemOcc reports 22.9 mIoU, which is identical to the 22.9 mIoU of OccLoff [44] listed in the same table. The text says the method 'improves mIoU from 20.1 (M-CONet) to 22.9' and claims 'substantial gains', but it does not acknowledge that this is a tie with the current state-of-the-art OccLoff. This is a fairness and completeness issue: the reader cannot assess whether the proposed method advances the state of the art when the leading comparison is omitted from the narrative. Please revise the discussion to state the tie explicitly and, if possible, add a statistical comparison (multiple seeds) to distinguish the methods.","section":"§4.2, Table 2"},{"comment":"The empirical claims rest on small differences (e.g., 27.97 vs. 28.21 for fusion strategy; 28.08 vs. 28.21 for SEAttention3D), but the paper reports no multiple-seed runs, error bars, or significance tests. Training hyperparameters (optimizer, learning rate, batch size, number of epochs, voxel size details) are not provided, and no code is released. For a deep-learning paper in a field where single-run differences of 0.1–0.2 mIoU are common, this lack of uncertainty quantification is load-bearing. At minimum, the authors should report the number of runs, mean and std, and provide the training configuration in an appendix.","section":"§4.1, §4.3"}],"minor_comments":[{"comment":"Typo: 'moudle' should be 'module' in the caption.","section":"Table 4"},{"comment":"The caption appears to contain duplicated 'Inception' text; Fig. 3 shows SEAttention3D, not inception modules. Please clean up the figure/caption.","section":"Fig. 3"},{"comment":"The height distribution plot has no axis labels or units. Without these, the justification for the chosen slice boundaries cannot be assessed.","section":"Fig. 1"},{"comment":"The title says 'nuScenes-Occupancy', but the text and the rest of the paper refer to 'nuScenes-OpenOccupancy'. Please unify the name.","section":"Table 2 title"},{"comment":"After reducing channels to one with a 3D convolution, A_global and A_local have shape B×1×X×Y×Z; the elementwise multiplication with F_global (B×C×X×Y×Z) requires broadcasting. Please state this explicitly.","section":"Eq. (5)"},{"comment":"The problem statement says labels are in [0,16], but the evaluation uses 16 semantic classes plus empty. Please clarify whether 0 is empty and classes are 1–16, and whether '16' is inclusive.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely a workshop-level submission. The main issue is not the architecture, which is coherent, but the framing of the results: the headline gain over M-CONet is inflated by an uncontrolled baseline, and the OpenOccupancy result is a tie with OccLoff. If the authors retrain M-CONet with their pipeline, report multiple seeds, and clarify the train-only nature of the slice boundaries, the contribution could be publishable. I would not recommend rejection, but the current manuscript is not acceptable without these changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read SliceSemOcc. The vertical-slice idea is sensible: giving the height axis its own attention and fusing global and local slices is a reasonable way to exploit the strong height structure in driving scenes. SEAttention3D, pooling over X and Y only, is a natural extension of SENet to voxels. The ablations support each component: adding the proposed modules to their own no-slice baseline moves mIoU from 26.98 to 28.21 on SurroundOcc, about +1.23. So the architecture does something.\n\nThe problem is the headline comparison. The paper claims a 14.2% relative gain over M-CONet (24.7 to 28.2), but their own baseline without any slicing already gets 26.98. That 2.3-point gap over the published M-CONet number is not explained by the proposed modules; it likely comes from the rest of the recipe—InverseMatrixVT3D projection, FPN, VoxelNet, and the focal+Lovasz+affinity losses. The paper never trains M-CONet under its own pipeline, so the central claim that vertical slicing significantly beats M-CONet is unsupported. What is supported is a modest internal gain over a baseline that is itself already stronger than M-CONet. The per-class small-object gains inherit the same confound. On OpenOccupancy they tie OccLoff at 22.9, which further undercuts the 'significant enhancement' framing.\n\nThe local slice boundaries are chosen from the height distribution of the same dataset they evaluate on. Without evidence they were set on the training split or that they transfer elsewhere, this is a mild test-set-informed design. It is not fatal but should be acknowledged, and the ablation with uniform bins (Table 4) helps, though the uniform bins are also arbitrary.\n\nNo code, no training hyperparameters, and no error bars. So we cannot rule out that the remaining 1.2-point gain is partly noise.\n\nWho should read it: people working on occupancy prediction who want to see a height-aware attention mechanism. It deserves a serious referee, but the evaluation needs major tightening before the claims are credible: retrain a proper M-CONet baseline, report multiple seeds, release code, and either validate the slice boundaries on other data or soften the claim.","headline":"Solid height-aware attention idea, but the reported gains over M-CONet are inflated by an uncontrolled baseline; the real internal gain is about 1.2 mIoU.","tokens_in":12774,"tokens_out":2970,"would_cite":false,"duration_ms":26519,"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":"SliceSemOcc claims that splitting voxel features into global and local vertical slices, fused by cross attention with per-height channel weights, improves 3D semantic occupancy mIoU by 14.2% on SurroundOcc and 13.9% on OpenOccupancy, with t","keywords":["3D semantic occupancy prediction","vertical slicing","channel attention","multimodal fusion","LiDAR-camera fusion","autonomous driving","nuScenes","voxel representation"],"falsifier":"A decisive check: recompute the slice boundaries from the training split's height histogram only (or use uniform bins) and evaluate on a dataset with a different height profile; if the mIoU advantage over M-CONet collapses or reverses, the vertical-slice benefit is dataset-specific. The paper's own Table 4 gives the comparison point: hand-picked bins score 28.21 mIoU versus 27.67 for uniform 1-meter bins, so the open question is whether that gap survives train-only boundary selection and cross-dataset transfer.","tokens_in":11761,"feed_emoji":"🚗","tokens_out":5502,"duration_ms":48293,"temperature":0.7,"pith_summary":"This paper tries to establish that 3D semantic occupancy prediction improves when the network treats the height axis explicitly instead of pooling over the whole voxel volume. It proposes two mechanisms: global and local vertical slices of the voxel features, fused by bidirectional cross attention, and a channel-attention variant that gives each height layer its own channel weights. If correct, the framework offers a modular way to get materially better voxel-level semantic labels, especially for small objects such as pedestrians, bicycles, and traffic cones that occupy narrow height bands. The claim is supported by mIoU gains on two nuScenes-based benchmarks and by ablations isolating each design choice.","feed_headline":"Vertical slices lift 3D occupancy accuracy by 14%","feed_subtitle":"Height-aware attention and global-local slices improve small-object classes like pedestrians and bicycles in driving scenes.","key_machinery":"The Vertical Slice Fusion (VSF) module and the SEAttention3D module. VSF splits the voxel volume along height into one global slice spanning [-5,3] m and six local slices concentrated where small objects sit, processes each with SEAttention3D, then fuses global and local features through bidirectional cross attention. SEAttention3D replaces SENet-style global pooling with pooling over only the X-Y plane, so it produces one channel-attention vector per height layer instead of one per volume, letting the network weight the same channel differently at different heights.","core_discovery":"The paper reports that processing 3D occupancy voxel features with vertical slices rather than treating the volume uniformly improves semantic occupancy prediction. On nuScenes-SurroundOcc, the full SliceSemOcc pipeline raises mean IoU from 24.7% for the M-CONet baseline to 28.2% (a 14.2% relative gain), with the largest per-class gains on small objects: bicycle +53.1%, motorcycle +37.2%, pedestrian +30.0%, barrier +21.3%, traffic cone +20.4%. On nuScenes-OpenOccupancy mIoU rises from 20.1% to 22.9% (a 13.9% relative gain). Ablations attribute the gains to the combination of global and local vertical slices, to choosing local slice boundaries from the class height distribution rather than un","pith_inferences":["The hand-picked slice boundaries are a dataset-specific choice; a learnable slicing mechanism could preserve the gains without retuning per dataset.","The height-aware channel weighting is not tied to occupancy prediction and could transfer to other voxel-based 3D perception tasks, such as LiDAR-camera 3D detection or scene completion.","The reported weakness on flat-ground classes suggests height alone cannot separate semantically different but geometrically similar ground classes; injecting LiDAR intensity or surface orientation would be a natural next test."],"forward_implications":["On nuScenes-SurroundOcc, SliceSemOcc raises mIoU from 24.7% (M-CONet) to 28.2%, a relative gain of 14.2%.","Gains are concentrated on small objects: barrier, bicycle, motorcycle, pedestrian, and traffic cone each improve by 20% to 53% relative.","On nuScenes-OpenOccupancy, mIoU rises from 20.1% to 22.9%, a relative gain of 13.9%.","Ablations show that height-distribution-based local bins beat uniform bins (28.21 vs 27.67 mIoU), and cross-attention fusion beats simple concatenation (28.21 vs 27.97 mIoU).","The vertical-slice fusion module costs about 0.6 GB of extra validation GPU memory for a 1.14-point mIoU gain."],"supporting_citations":[{"why":"Supplies the OpenOccupancy benchmark, the M-CONet multimodal baseline, and the ground-truth occupancy labels used in the second evaluation.","marker":"[35]"},{"why":"Supplies the SurroundOcc dataset and the object height distribution (Fig. 1) that motivates the choice of global and local vertical slice boundaries.","marker":"[38]"},{"why":"Provides the projection-matrix view-to-voxel transformation that turns multi-view camera features into 3D voxel features in the camera branch.","marker":"[24]"},{"why":"Provides the VoxelNet 3D backbone that encodes voxelized LiDAR sweeps into point-cloud voxel features.","marker":"[46]"},{"why":"The OccFusion multimodal occupancy fusion framework that the paper compares against and positions its height-aware fusion relative to.","marker":"[25]"},{"why":"Co-Occ is a strong multimodal occupancy baseline in both result tables, used to situate the reported mIoU gains.","marker":"[27]"},{"why":"Supplies MonoScene, a camera-only baseline, and the affinity losses Lgeo and Lsem used in the total training objective.","marker":"[4]"}],"fun_headline_variants":["Vertical slices cut through to 14% better 3D occupancy","Height-aware slices improve 3D occupancy by 14%","Slicing voxels vertically improves small-object occupancy","Global-local slices sharpen 3D occupancy by 14%","Vertical slicing raises 3D occupancy mIoU 14%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The local slice boundaries were picked by looking at the height distribution of objects in the same dataset the method is then evaluated on, so the reported gains may partly reflect tuning to that dataset's statistics rather than a general principle.","fun_headline_variants_meta":{"raw":{"variants":["Vertical slices cut through to 14% better 3D occupancy","Height-aware slices improve 3D occupancy by 14%","Slicing voxels vertically improves small-object occupancy","Global-local slices sharpen 3D occupancy by 14%","Vertical slicing raises 3D occupancy mIoU 14%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2736,"prompt_tokens":806,"completion_tokens":1930,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1858}},"tokens_in":550,"tokens_out":1930,"duration_ms":13404,"temperature":1.0,"reasoning_tokens":1858,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:28:01.526801+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check: recompute the slice boundaries from the training split's height histogram only (or use uniform bins) and evaluate on a dataset with a different height profile; if the mIoU advantage over M-CONet collapses or reverses, the vertical-slice benefit is dataset-specific. The paper's own Table 4 gives the comparison point: hand-picked bins score 28.21 mIoU versus 27.67 for uniform 1-meter bins, so the open question is whether that gap survives train-only boundary selection and cross-dataset transfer.","supporting_citations":[{"cited_title":"InverseMatrixVT3D: An Efficient Projection Matrix-Based Approach for 3D Occupancy Prediction","cited_arxiv_id":"2401.12422","evidence_quote":"Provides the projection-matrix view-to-voxel transformation that turns multi-view camera features into 3D voxel features in the camera branch."},{"cited_title":"IEEE Robotics Autom","cited_arxiv_id":null,"evidence_quote":"Co-Occ is a strong multimodal occupancy baseline in both result tables, used to situate the reported mIoU gains."}],"review_version":1}