{"id":"de3f218e-f48e-4d3c-8ba6-b84bfc64e414","arxiv_id":"2501.16684","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A camera-only indoor 3D semantic occupancy model using paired-horizontal-plane features achieves 15.45% mIoU on EmbodiedScan, best among evaluated RGB baselines.","lead":"SliceOcc introduces a vertical slice representation for indoor 3D semantic occupancy prediction, slicing the scene along the height axis into paired plane features and fusing them with attention. It reports state-of-the-art mIoU (15.45%) on EmbodiedScan among RGB camera-based models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is undercut by unmatched input views: OccNet and SurroundOcc are not re-run with SliceOcc's 20-view protocol, and the paper's own Fig. 4 shows view count is a strong performance lever.","rationale":"The reader's verdict of CONDITIONAL is appropriate, and the reader's rationale already mentions the partial fair-comparison setup. However, the reader's stated weakest assumption is the horizontal-slice geometric prior, whereas I see the unmatched input-view protocol as the more directly load-bearing threat to the paper's central empirical claim. The horizontal-slice assumption could fail on slanted geometry, but that failure would manifest as a drop in absolute performance; the paper's SOTA claim would still be false if a view-matched baseline outperforms SliceOcc. The paper deserves credit for releasing code, running ablations on slice number, resolution, layer number, and components, and re-implementing EmbodiedScan with the same view count. Those elements support the method's internal consistency and its advantage over one strong baseline. But the headline 'state-of-the-art among RGB camera-based models' requires all listed RGB baselines to be evaluated under matched conditions. The proposed test is straightforward: re-run OccNet and SurroundOcc with 20 views. If they remain below SliceOcc, the conditional acceptance can be upgraded; if not, the SOTA claim should be weakened to 'competitive with RGB baselines under matched view counts.' Thus the reader's conditional verdict stands unchanged, but for a more sharply identified reason.","tokens_in":10719,"tokens_out":3849,"duration_ms":40213,"concrete_test":"Re-run OccNet and SurroundOcc under exactly SliceOcc's protocol: 20 input views, the same ResNet-50 backbone, the same EmbodiedScan training split, and the same training schedule and loss. Evaluate on the same 389-scene test set. If either baseline reaches or exceeds 15.45% mIoU, the SOTA claim is unsupported. As a complementary check, evaluate SliceOcc with the original 10-view protocol to quantify the view-count contribution to the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SliceOcc is state-of-the-art among RGB camera-based models rests on Table I. There, SliceOcc uses 20 input views, and the only view-matched baseline is EmbodiedScan dagger, which is re-implemented with 20 views and scores 14.52. OccNet (8.07) and SurroundOcc (9.10) are reported at their original settings, which in their source papers use fewer cameras (typically 6 or 10). The paper's Fig. 4 explicitly demonstrates that SliceOcc's mIoU increases with the number of input views, so view count is a known confound. Without view-matched re-runs of OccNet and SurroundOcc, the 15.45% mIoU does not establish a fair SOTA comparison. The horizontal-slice assumption is a plausible inductive bias, but the empirical superiority claim is the load-bearing part of the paper, and it is currently supported by only one view-matched baseline. This is not an allegation of dishonesty; it is an incomplete control in the experimental design.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"SliceOcc proposes a camera-only indoor 3D semantic occupancy prediction method based on a vertical slice representation. The scene is divided along the vertical axis into S slices, each represented by floor and ceiling plane queries; sliced spatial cross-attention and planar cross-attention update these queries from multi-view images, and voxel features are reconstructed by interpolating between neighboring slice planes. On the EmbodiedScan benchmark, SliceOcc reports 15.45% mIoU across 81 categories, outperforming reported RGB baselines and approaching the depth-based Mink-ResNet34. The paper includes ablations of slice number, slice resolution, layer number, input view count, and the main attention components.","tokens_in":10920,"tokens_out":7535,"duration_ms":72867,"significance":"If the empirical results hold under a properly controlled comparison, the vertical slice representation is a plausible and effective inductive bias for indoor occupancy prediction, where horizontal surfaces and vertical consistency are common. The paper has several strengths: it releases code, re-implements a baseline (EmbodiedScan†) with the same input view count, explicitly studies the effect of input views, and provides component-wise ablations. The central assumption about indoor geometry is stated transparently in the introduction. However, the headline state-of-the-art claim currently rests on a comparison in which only one RGB baseline is view-matched, and no variance estimates are reported; these issues must be addressed before the main claim is convincing.","major_comments":[{"comment":"The state-of-the-art claim is undercut by an input-view mismatch. SliceOcc uses N=20 input views (Section IV-B), and Table I reports OccNet (8.07) and SurroundOcc (9.10) without stating that they were re-run with 20 views; the only baseline described as view-matched is EmbodiedScan† (14.52). Because Fig. 4 shows that increasing the input views from 10 to 20 improves SliceOcc's mIoU, the reported gap over OccNet and SurroundOcc may be partly explained by the richer input rather than by the vertical slice representation. Please re-run OccNet and SurroundOcc under the same 20-view protocol (and ideally the same training schedule and image resolution), or clearly qualify the claim as based on a protocol-inconsistent comparison.","section":"Table I; Fig. 4"},{"comment":"All quantitative results are single-run numbers without error bars or multiple seeds. The main improvement over the view-matched baseline is 0.93 mIoU (15.45 vs 14.52), and many category-level differences are within a few points; without variance estimates these differences cannot be distinguished from training noise. Please report the mean and standard deviation over at least three independent runs for the main comparison and for the core ablations (slice number, slice resolution, and component ablations).","section":"Tables I–V"},{"comment":"The paper motivates the vertical slice representation by contrasting it with planar representations (BEV and TPV), but it never evaluates a TPV-based method under the same experimental protocol. The slice-number-1 single-plane variant is not a full BEV model, and the baselines OccNet and SurroundOcc are not TPV methods. Thus the contribution claim that vertical slices are superior to prior planar representations is not directly supported. Adding a view-matched TPV-style baseline (e.g., a TPVFormer variant on EmbodiedScan) or explicitly limiting the claim to the evaluated baselines would strengthen the paper.","section":"Section IV-E; Table V"}],"minor_comments":[{"comment":"The voxel-feature reconstruction step is described only as \"interpolating its neighboring ceiling slice feature Qc_s and floor slice feature Qf_s using tri-linear methods.\" Since each slice feature is a W×L planar map, the operation is bilinear interpolation within each plane followed by linear interpolation along height; the term \"tri-linear\" is imprecise. The text should also specify how points near the bottom and top scene boundaries are handled and how the interpolation weight is normalized by slice height.","section":"Section III-C"},{"comment":"The re-implementation of EmbodiedScan† is described only as \"optimized with the same input view number as ours.\" To make the view-matched baseline reproducible and trustworthy, please report the other training settings used for this re-implementation (e.g., backbone, image resolution, training epochs, data augmentation, and loss function).","section":"Section IV-C"},{"comment":"The motivating assumption that indoor objects \"mostly have horizontal surfaces and consistent vertical orientation\" is stated clearly, but the paper does not discuss or analyze failure cases such as slanted walls, stairs, ramps, or overhanging objects. A brief limitation discussion or a small qualitative analysis of such scenes would make the scope of the representation claim more precise.","section":"Section I"},{"comment":"There is an inconsistency in naming: the text refers to \"OCCNet\" while Table I uses \"OccNet\"; please unify the terminology. In addition, Eq. (8) uses β ∈ {0, 1/2} without stating which value corresponds to the floor query and which to the ceiling query.","section":"Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is a well-structured empirical architectural study with code release and a thoughtful ablation setup. The main obstacle is not the method itself but the experimental protocol: the headline SOTA claim needs view-matched re-runs of OccNet and SurroundOcc, and ideally variance estimates. I see no signs of misconduct; the issue is an incomplete control. If the authors can provide the matched baselines or carefully revise the claim, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: SliceOcc is a genuinely incremental but useful architectural idea—vertical slices instead of one BEV plane or three TPV planes—and the authors back it with clean ablations and a code release. But the headline SOTA claim is not fully established, because Table I compares SliceOcc at 20 input views against OccNet and SurroundOcc at their original, lower view counts, and their own Fig. 4 shows view count is a strong performance lever. The only view-matched baseline is EmbodiedScan†, and the gap there is 0.93 mIoU (14.52 vs 15.45). That is a modest, plausible gain, but it is not a decisive win over a fair baseline set.\n\nWhat's new: the vertical slice representation, splitting the scene into S horizontal slab pairs and using deformable attention to query image features onto the nearest pair of planes, then interpolating back to voxels. This is a natural extension of BEV/TPV and the paper shows it helps: slice number ablations (2→8→16) improve mIoU monotonically, and the component ablation (Table V) attributes clear gains to SSCA and PCA. The method is specified in enough detail that reimplementation is feasible, and they release code.\n\nSoft spots, in order of severity. First, the unmatched baselines. OccNet and SurroundOcc are not re-run with SliceOcc's 20-view protocol, so the comparison is confounded. The authors should re-run them, or at least provide a view-matched ablation. Second, no multiple seeds or error bars, so we don't know if 0.93 mIoU is noise. Third, the domain assumption about horizontal surfaces is reasonable for indoor scenes, but it could limit generalization to slanted/overhanging geometry; that is a limitation worth stating explicitly rather than a flaw in the method.\n\nOverall, the empirical evidence supports the representation's value, but the \"state-of-the-art\" claim needs stronger controls. This is a borderline-accept paper, not a clear accept. It deserves a serious referee slot: the idea is sensible, the experiments are mostly well-designed, and the subfield (indoor occupancy prediction) needs more camera-only baselines.\n\nRecommendation: send to peer review with a request for view-matched baseline comparisons and repeated runs. If the authors can close that gap, the paper is solid.\n\nBest,\n[Your name]","headline":"A useful incremental representation with a SOTA claim undercut by unmatched input views; deserves review with view-matched baselines.","tokens_in":11497,"tokens_out":2247,"would_cite":true,"duration_ms":21193,"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":"Vertical slice representation enables RGB-based indoor 3D occupancy prediction at 15.45% mIoU.","keywords":["3D semantic occupancy prediction","vertical slice representation","indoor scene understanding","multi-view RGB images","deformable cross-attention","EmbodiedScan benchmark","planar scene representation","occupancy prediction"],"falsifier":"Evaluate SliceOcc on an indoor subset containing many staircases, ramps, sloped ceilings, or leaning objects and compare its per-class mIoU against a full-voxel or tri-perspective baseline; if the slice-based model degrades disproportionately on those classes while the baseline does not, the slice representation is the limiting factor.","tokens_in":10503,"feed_emoji":"🏠","tokens_out":6128,"duration_ms":50681,"temperature":0.7,"pith_summary":"The paper sets out to show that for indoor scenes, a vertical slice representation—dividing the 3D space into horizontal slabs and encoding each slab by a floor plane and a ceiling plane—is a more effective planar representation for 3D semantic occupancy prediction than BEV or tri-perspective views. On this representation the authors build SliceOcc, an RGB-camera transformer that extracts slice features with two cross-attention blocks and interpolates them into voxel predictions. On the EmbodiedScan benchmark, SliceOcc reaches 15.45% mean IoU over 81 classes, the best among camera-only models and close to depth-based models. The implication is that dense, heavily occluded indoor scenes can be understood from images alone, without explicit depth input.","feed_headline":"Vertical slices lift RGB indoor occupancy to 15.45% mIoU","feed_subtitle":"A camera-only model beats BEV and TPV approaches on EmbodiedScan's 81 indoor classes.","key_machinery":"The vertical slice representation is the central object: the scene is divided uniformly along the vertical axis into $S$ parts, and each part is represented by a floor-plane query and a ceiling-plane query. Any spatial point is reconstructed by projecting its feature onto the nearest plane pair and interpolating. Two deformable-attention blocks carry the method: SSCA samples pillar-shaped 3D reference points around each plane, projects them into the input images, and applies multi-scale deformable attention to pull image features onto the slice; PCA applies deformable attention between the floor and ceiling queries of the same slice to enforce vertical consistency. Slice queries are initialized by grid-sampling image features at 3D anchor positions and adding learnable height embeddings for each slice level.","core_discovery":"SliceOcc's central claim is that full 3D semantic occupancy of an indoor scene can be recovered from multi-view RGB images by representing the scene as $S$ stacked horizontal slices, each captured by a pair of parallel plane feature maps (floor and ceiling), and by trilinearly interpolating between adjacent planes to form voxel features. With 16 slices at $40\\times40$ resolution and 20 input views, the model attains 15.45% mIoU across 81 classes on EmbodiedScan, beating the RGB baselines OccNet, SurroundOcc, and EmbodiedScan (including a re-implemented 20-view EmbodiedScan at 14.52%) and matching the depth-input Mink-ResNet34. The authors attribute the gain to the slice representation itself and to the two attention blocks, sliced spatial cross-attention (SSCA) and planar cross-attention (PCA), which let each plane query exchange information with image features and with its paired plane.","pith_inferences":["The paper does not test this, but if the vertical slice representation is responsible for the gain, its performance should drop on scenes dominated by slanted surfaces such as staircases and ramps; an evaluation on such a subset would isolate the representation's contribution from the attention modules' contribution.","A natural extension the paper leaves implicit is adaptive slice placement: concentrating more slices near the floor, where most indoor objects sit, could improve mIoU at the same computational budget.","Beyond the paper, the slicing argument is tied to indoor geometry; applying the same method outdoors along the gravity axis would likely yield smaller gains because outdoor scenes have fewer consistent horizontal surfaces.","The paper's comparison to depth-based models is single-benchmark; whether the gap stays closed on unseen scenes or irregular object shapes is an open question the paper does not address."],"forward_implications":["Camera-only indoor occupancy prediction reaches accuracy comparable to depth-based models, so depth sensors may not be necessary for dense indoor scene understanding.","Performance improves monotonically with the number of slices in the tested range (12.13% at $S=2$, 14.57% at $S=8$, 15.45% at $S=16$), showing that vertical resolution is a key capacity lever.","Ablations show SSCA and PCA are complementary: 15.12% with SSCA alone, 14.70% with PCA alone, 15.36% combined, and 15.45% with the height query added.","The method generalizes across the three indoor datasets in the EmbodiedScan benchmark (ScanNet, Matterport3D, 3RScan), since training and testing draw on all of them."],"supporting_citations":[{"why":"Supplies the EmbodiedScan benchmark, occupancy labels, and the main RGB baseline (EmbodiedScan/EmbodiedScan†) that SliceOcc must beat.","marker":"[27]"},{"why":"SurroundOcc is the tri-perspective-view RGB baseline that SliceOcc outperforms.","marker":"[3]"},{"why":"OccNet is the scene-as-occupancy RGB baseline that SliceOcc outperforms.","marker":"[35]"},{"why":"MonoScene provides the Scene-Class Affinity Loss (geometric and semantic components) used in training.","marker":"[31]"},{"why":"Deformable DETR supplies the deformable attention operator used by both SSCA and PCA.","marker":"[30]"},{"why":"TPVFormer introduces the tri-perspective view planar representation that SliceOcc contrasts with.","marker":"[1]"},{"why":"ScanNet provides the annotated RGB-D scenes used as part of the EmbodiedScan training and test data.","marker":"[32]"},{"why":"Matterport3D provides indoor building scenes used in the benchmark.","marker":"[33]"},{"why":"3RScan provides indoor scans used in the benchmark.","marker":"[34]"}],"fun_headline_variants":["Vertical slice representation beats BEV and TPV for indoor occupancy","Camera-only 3D occupancy: vertical slices top EmbodiedScan at 15.45% mIoU","Stacked planes let RGB cameras predict indoor occupancy best yet","Indoor 3D occupancy: vertical slices outdo planar baselines on EmbodiedScan"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that indoor objects mostly have horizontal surfaces and consistent vertical orientation, so a fixed set of horizontal slices with interpolation between plane features can faithfully reconstruct the full 3D occupancy.","fun_headline_variants_meta":{"raw":{"variants":["Vertical slice representation beats BEV and TPV for indoor occupancy","Camera-only 3D occupancy: vertical slices top EmbodiedScan at 15.45% mIoU","Stacked planes let RGB cameras predict indoor occupancy best yet","Indoor 3D occupancy: vertical slices outdo planar baselines on EmbodiedScan"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001312,"raw_usage":{"total_tokens":5386,"prompt_tokens":1025,"completion_tokens":4361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":4275}},"tokens_in":641,"tokens_out":4361,"duration_ms":27152,"temperature":1.0,"reasoning_tokens":4275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T11:25:55.653531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate SliceOcc on an indoor subset containing many staircases, ramps, sloped ceilings, or leaning objects and compare its per-class mIoU against a full-voxel or tri-perspective baseline; if the slice-based model degrades disproportionately on those classes while the baseline does not, the slice representation is the limiting factor.","supporting_citations":[{"cited_title":"Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai,","cited_arxiv_id":null,"evidence_quote":"Supplies the EmbodiedScan benchmark, occupancy labels, and the main RGB baseline (EmbodiedScan/EmbodiedScan†) that SliceOcc must beat."},{"cited_title":"Scene as occupancy,","cited_arxiv_id":null,"evidence_quote":"OccNet is the scene-as-occupancy RGB baseline that SliceOcc outperforms."},{"cited_title":"Monoscene: Monocular 3d semantic scene completion,","cited_arxiv_id":null,"evidence_quote":"MonoScene provides the Scene-Class Affinity Loss (geometric and semantic components) used in training."},{"cited_title":"Deformable detr: Deformable transformers for end-to-end object detection,","cited_arxiv_id":null,"evidence_quote":"Deformable DETR supplies the deformable attention operator used by both SSCA and PCA."},{"cited_title":"Tri-perspective view for vision-based 3d semantic occupancy prediction,","cited_arxiv_id":null,"evidence_quote":"TPVFormer introduces the tri-perspective view planar representation that SliceOcc contrasts with."},{"cited_title":"Rio: 3d object instance re-localization in changing indoor environments,","cited_arxiv_id":null,"evidence_quote":"3RScan provides indoor scans used in the benchmark."}],"review_version":1}