{"id":"eb3b07c2-6db8-4e16-a0c9-0ab8ac6f3688","arxiv_id":"2412.07163","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fast occupancy network lifts BEV features to voxels with deformable 2D convolution and a partial voxel FPN, beating OccNet's mIoU on OpenOcc at roughly one third the latency.","lead":"A team at AMD shows that a camera-based 3D occupancy network for self-driving cars can run about three times faster by lifting bird's-eye-view features into 3D with a deformable 2D convolution instead of expensive 3D attention. Their model reports higher accuracy than the OccNet baseline on the OpenOcc benchmark, and a training-only auxiliary segmentation loss adds no cost at inference.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3's SemanticKITTI mIoU for the proposed method is internally inconsistent: the per-class values average 12.27, not 12.44, reversing the claimed 0.16% edge over OccFormer.","rationale":"The reader's weakest_assumption is that BEV features encode enough vertical information for the 2D deformable-convolution lifting to recover 3D voxel features. That is a generalization risk, not an internal inconsistency, and it is partially supported by Table 5's ablations and by OccNet's own BEVNet result. The Table 3 arithmetic issue is more load-bearing for the stated claim because it is an internal, checkable discrepancy in a benchmark explicitly used to argue for 'consistently outperforms.' The paper's own mIoU convention can be inferred from the baseline rows: for TPVFormer, the mean of the 19 non-Road columns is 11.26, matching the published 11.26, and for OccFormer it is 12.32, matching 12.32. The same calculation on the proposed method's row gives 12.27, not 12.44. If confirmed, the proposed method is not more accurate than OccFormer on SemanticKITTI, although it is still faster and still outperforms on OpenOcc. This does not by itself invalidate the core technical contribution, but it does puncture the unqualified 'consistently outperforms' claim. The reader already reached CONDITIONAL and flagged the arithmetic issue in the rationale, so the verdict does not need to move; it should explicitly include verification of Table 3 as a condition. I therefore mark agreement as partial: the reader identified the same issue but did not make it the weakest assumption.","tokens_in":14665,"tokens_out":10237,"duration_ms":89631,"concrete_test":"Recompute the mean of the 19 non-Road IoU columns in Table 3's 'Ours' row using the same convention that reproduces TPVFormer's 11.26 and OccFormer's 12.32. If the mean is 12.27 rather than 12.44, the paper's claimed 0.16% advantage over OccFormer reverses to a small deficit, and the 'consistently outperforms' claim loses its SemanticKITTI support. Request the authors' raw per-class scores or test-set predictions to determine which value is correct.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that the method 'consistently outperforms existing methods in both accuracy and inference speed.' The OpenOcc results in Tables 1 and 4 are internally coherent, but the only SemanticKITTI accuracy table, Table 3, contains a load-bearing arithmetic inconsistency. In Table 3, the reported mIoU is not the mean of all 20 listed category columns; for the published baselines it matches the mean of the 19 non-Road columns (TPVFormer: 11.26; OccFormer: 12.32). Applying that same convention to the 'Ours' row, the 19 non-Road IoU values (55.90, 29.60, 32.80, 11.40, 15.00, 21.20, 2.20, 1.10, 1.60, 4.40, 15.90, 2.20, 19.80, 1.30, 1.10, 0.50, 11.70, 3.30, 2.10) average 12.27, not the reported 12.44. Section 4.4 uses the 12.44 value to claim a 0.16% improvement over OccFormer; under the table's own protocol, the proposed method would be 12.27, slightly below OccFormer's 12.32. Because 'consistently outperforms' is the headline claim and Table 3 is the only SemanticKITTI evidence, this discrepancy is directly load-bearing: if it stands, the consistency claim is false and only the OpenOcc benchmark remains. The introduction's 'we prove' statement is also unsupported by any proof, but the empirical central claim does not depend on it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Fast Occupancy Network, a camera-based 3D semantic occupancy model whose main components are a BEV lifter based on 2D deformable convolution, a Partial Voxel FPN, and a perspective-view auxiliary segmentation loss used only at training time. On the OpenOcc/nuScenes benchmark it reports 21.12 mIoU with ResNet50 at 1.22x relative latency versus BEVNet, compared with OccNet at 19.48 mIoU and 3.43x latency, and it reports 27.22 mIoU with ResNet101-DCN. On SemanticKITTI it reports 12.44 mIoU and claims a 0.16% advantage over OccFormer. The paper argues that BEV features already contain enough vertical information for a 2D deformable convolution, applied on the BEV plane and reshaped, to replace costly 3D deformable attention.","tokens_in":14972,"tokens_out":10548,"duration_ms":87237,"significance":"If the OpenOcc results are reproducible, this is a useful empirical contribution: it shows that a deformable 2D convolution can lift BEV features to 3D voxel features with a large latency reduction and improved accuracy on that benchmark. The ablations in Tables 5-7 are a genuine strength because they isolate the contributions of the lifting module, PV supervision, visible mask, and Partial Voxel FPN, and they report latency proportions. The main caveat is that the SemanticKITTI evidence contains an arithmetic inconsistency in the reported mIoU that invalidates the paper's cross-dataset 'consistently outperforms' claim as currently stated. No code or machine-checked proofs are provided, so reproducibility rests on the internal consistency of the tables and the reported experimental protocol.","major_comments":[{"comment":"The reported SemanticKITTI mIoU of 12.44 for the proposed method is not the mean of the 19 listed class IoUs. Under the same convention used for every other row in Table 3 (including OccFormer, whose listed values average to 12.32), the 'Ours' row values sum to 233.1 and average to 12.27, which is below OccFormer. Since Section 4.4 explicitly claims a 0.16% improvement over OccFormer and Table 4 repeats the 12.44 value, the central 'consistently outperforms existing methods' claim currently rests on an arithmetic error. Please correct either the per-class values or the reported mIoU, and revise the text and Table 4 accordingly.","section":"Table 3; Section 4.4; Table 4"}],"minor_comments":[{"comment":"The BEVNet† Res101-DCN row reports mIoU 24.62, but the 16 listed class IoUs average to approximately 23.78. This is a baseline row and does not weaken the paper's positive OpenOcc comparison, but the main comparison table should be internally consistent and the row should be corrected or footnoted.","section":"Table 1"},{"comment":"The statement 'We prove that this lifting method is able to recover voxel information from the BEV feature' is not supported by any proof or formal argument anywhere in the manuscript; please replace 'prove' with an empirical claim or provide an actual proof.","section":"Section 1, last paragraph"},{"comment":"The claim that 2D and 3D deformable cross-attention differ only in the number of queries and in whether height is collapsed at the end is imprecise, because 3D deformable attention also uses 3D reference points; rephrase to describe the actual difference accurately.","section":"Section 3.2"},{"comment":"The perspective-view branch is 'cost-free' only at inference time; it adds training-time computation and parameters, so the claim should be qualified as 'inference-cost-free.'","section":"Abstract; Section 3.2"},{"comment":"The sentence 'Here, WP Vand Wvox are represented as the weights of the weights of the PV loss and the voxel loss' contains duplicated wording and should be rewritten.","section":"Equation (3) and surrounding text"},{"comment":"The SemanticKITTI description says there are 21 semantic classes (19 semantics, 1 free, 1 unknown), but Table 3 reports only 19 per-class columns; please clarify which classes are included in the reported mIoU.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The SemanticKITTI mIoU discrepancy is the kind of issue that should be checked against the raw per-class predictions before publication. The OpenOcc evidence is internally coherent and the ablations are informative, so I do not see grounds for rejection at this stage, but the authors must correct the arithmetic in Table 3 and re-examine the affected cross-dataset claims in Sections 4.4 and Table 4. The editor may also wish to request the evaluation scripts or per-class results for SemanticKITTI to verify the corrected numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing: the OpenOcc numbers are coherent, and the deformable-2D-conv lifter with partial voxel FPN is a genuinely simple efficiency idea. But the SemanticKITTI result does not survive contact with its own table. The reported mIoU for 'Ours' is 12.44; the mean of the 19 per-class IoUs in the same row is 12.27. The paper's baselines use the same 19-class mean, so under the paper's own protocol Ours is actually below OccFormer's 12.32. That reverses the 0.16% edge claimed in Section 4.4 and undercuts the abstract's 'consistently outperforms.' This isn't a rounding issue; it's a one-point swing in the opposite direction. The authors need to fix or explain this before the accuracy claims can be taken at face value.\n\nThere is also a likely copy-paste error in Table 1: the BEVNet† Res101-DCN row's listed class IoUs average to 23.78, not the reported 24.62. For the other rows the average matches, so this looks like a transcription slip rather than systematic gaming. Still, it makes the tables feel under-audited. And the intro's 'we prove that this lifting method...' is not backed by a proof anywhere; a more careful wording like 'we show empirically' would be appropriate. No code or error bars are provided, which limits how much confidence one can place in the efficiency numbers.\n\nWhat the paper does well: the design is simple, the ablations in Tables 5-7 are informative, and the OpenOcc main comparison is internally consistent. If the reported latencies hold, a ~3x speedup at better mIoU on OpenOcc is a useful contribution. The paper also clearly identifies the bottleneck and explains why 2D deformable convolution is a sensible lifting mechanism. The PV supervision with projected sparse labels is a nice free-lunch trick.\n\nThe problems are concentrated in the reporting, not the architecture. A careful revision that corrects the SemanticKITTI arithmetic, fixes the Table 1 row, and tones down the 'prove' claim could make this a solid conference paper. But as is, the central 'consistent outperformance' claim is unsupported.\n\nI'd send it to peer review because the core idea and OpenOcc evidence deserve scrutiny. But I'd flag the SemanticKITTI inconsistency as a major revision item, not a minor typo.","headline":"OpenOcc results hang together, but the SemanticKITTI table's own arithmetic reverses the paper's headline accuracy claim.","tokens_in":15642,"tokens_out":10146,"would_cite":false,"duration_ms":74931,"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":"The paper claims that 3D occupancy prediction from cameras can be made roughly three times faster without sacrificing accuracy by lifting BEV features to voxels with a 2D deformable convolution instead of 3D attention.","keywords":["occupancy prediction","BEV lifting","deformable convolution","voxel feature pyramid","semantic scene completion","autonomous driving","perspective view supervision","inference speed"],"falsifier":"An experiment that would settle the claim: keep the same occupancy decoder but replace the attention-based BEV construction with a depth-splatting (LSS-style) BEV builder on the OpenOcc benchmark, then compare the deformable-conv lifter against an MLP lifter. If the deformable-conv advantage shrinks to near zero, the paper's premise that BEV features contain recoverable vertical information is specific to attention-based BEV construction.","tokens_in":14390,"feed_emoji":"🚗","tokens_out":10975,"duration_ms":93599,"temperature":0.7,"pith_summary":"The paper tries to establish that expensive 3D attention is not needed to turn camera-based BEV features into 3D occupancy predictions. It claims that a 2D deformable convolution applied on the BEV plane and then reshaped into height layers lifts BEV features to voxel features as accurately as 3D deformable attention, at a fraction of the cost. Two further additions, a Partial Voxel FPN that fuses multi-scale features cheaply on the horizontal plane and a train-only perspective-view segmentation loss, raise accuracy with nearly zero inference cost. On the OpenOcc benchmark the complete model surpasses OccNet by 1.7% mIoU (21.12 vs. 19.48) with a ResNet50 backbone while running at about one third of OccNet's latency. A sympathetic reader would care because this suggests camera-only occupancy prediction can become fast enough for real-time driving systems without giving up accuracy.","feed_headline":"Occupancy network runs 3x faster, beats OccNet by 1.7% mIoU","feed_subtitle":"BEV features carry enough height information that a flat convolution lifts them to voxels at a fraction of the cost.","key_machinery":"The load-bearing mechanism is the BEV lifter: a deformable 2D convolutional layer that expands the BEV feature channels and then uses a reshape to split them into the height dimension, converting a $H \\times W \\times C$ BEV map into a $H \\times W \\times Z \\times C'$ voxel feature volume. Its dynamic offsets give an adaptive receptive field, which the paper says matters because objects at different heights have different geometries. The second mechanism is the Partial Voxel FPN, which downsamples only the planar (xy) dimensions at each scale, keeps half the original features untouched, upsamples and adds the downsampled half, and applies one small 3D convolution at the coarsest scale ($50 \\times 50 \\times 4$) to restore height interactions. The third is the perspective-view supervision branch: a U-Net-like decoder attached to the image backbone, supervised by sparse labels made by projecting LiDAR points into each camera view, trained with focal loss and removed at inference.","core_discovery":"The central claim is that BEV features already carry enough vertical information that a deformable 2D convolution over the BEV plane, followed by a channel-to-height reshape, produces voxel features comparable to 3D deformable cross-attention. The paper argues that the difference between 2D deformable attention and 3D deformable attention is mainly the number of queries and whether the height dimension is collapsed at the end, while the reference features are the same. It then packages this lifter with a Partial Voxel FPN that handles the height dimension cheaply, downsampling only the xy plane on half of the features and applying a small 3D convolution at the coarsest scale, and with a perspective-view auxiliary segmentation branch used only in training. The combined result, as reported, is a model that beats the previous state of the art on OpenOcc in both mIoU and latency, and also improves the LiDAR-segmentation-style evaluation derived from the occupancy prediction.","pith_inferences":["The paper's logic implies that the 3D query stacks in other occupancy models perform redundant work if BEV features already encode height; a testable consequence is that replacing those stacks with this lifter on other BEV backbones would preserve most accuracy.","The lifter's success may depend on how the BEV is built: attention-based view transformers may imprint vertical structure, while depth-splatting (LSS-style) builders may not, so the trade-off should be re-measured on a depth-splatting BEV feature.","The perspective-view supervision idea could transfer to other BEV tasks with long gradient paths, such as online HD map construction, where a cheap image-space auxiliary loss might give a similar free gain.","Latency ratios are measured on specific accelerator hardware; on other GPUs the relative speedup could differ because deformable convolution and the small 3D convolution have different kernel optimization profiles."],"forward_implications":["Any BEV perception model can be converted into an occupancy model by attaching this lifter and head; the paper demonstrates the idea on a BEVFormer-style builder, with about 22% extra latency over the plain BEVNet on ResNet50.","The cost of building voxel features no longer scales with the number of 3D queries, so larger occupancy ranges or finer voxel grids become more affordable.","The perspective-view supervision adds roughly 1.1% mIoU for free at inference, suggesting image-space supervision can compensate for the long gradient path through BEV transformation.","The Partial Voxel FPN reaches mIoU comparable to a full 3D FPN while using about one quarter of the latency, indicating most height interactions can be handled with mostly-2D operations."],"supporting_citations":[{"why":"Supplies the BEVNet baseline, the 3D deformable attention alternative, and the OpenOcc benchmark.","marker":"[39]"},{"why":"Supplies the BEVFormer view transformer that produces the BEV features the lifter operates on.","marker":"[19]"},{"why":"Origin of the deformable attention/convolution formulation whose 2D version the paper adapts as the lifter.","marker":"[48]"},{"why":"Inspiration for multi-scale voxel FPN fusion, which the Partial Voxel FPN makes cheaper.","marker":"[20]"},{"why":"State-of-the-art occupancy baseline the method compares against on OpenOcc and SemanticKITTI.","marker":"[12]"},{"why":"OccFormer, the SemanticKITTI baseline the method claims to surpass in mIoU at about half the latency.","marker":"[47]"},{"why":"The LSS view transformation pipeline cited as the alternative way to build BEV features.","marker":"[32]"},{"why":"BEVDet, whose scene-completion head forms the BEV-based occupancy baseline used by OccNet and this paper.","marker":"[10]"}],"fun_headline_variants":["Flat 2D conv lifts BEV to voxels, 3x faster","Occupancy nets 3x faster, beat OccNet by 1.7% mIoU","BEV height hidden in features: 2D lift cuts cost 3x","2D lift from BEV to voxels: 3x speedup, SOTA mIoU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that BEV features already encode enough vertical information that a 2D deformable convolution on the BEV plane, reshaped into height layers, can recover 3D voxel features as accurately as 3D deformable attention; if this transfer fails for other BEV builders or resolutions, the reported speed-accuracy trade-off collapses.","fun_headline_variants_meta":{"raw":{"variants":["Flat 2D conv lifts BEV to voxels, 3x faster","Occupancy nets 3x faster, beat OccNet by 1.7% mIoU","BEV height hidden in features: 2D lift cuts cost 3x","2D lift from BEV to voxels: 3x speedup, SOTA mIoU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000783,"raw_usage":{"total_tokens":3491,"prompt_tokens":1012,"completion_tokens":2479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":2383}},"tokens_in":628,"tokens_out":2479,"duration_ms":19002,"temperature":1.0,"reasoning_tokens":2383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:05:13.658848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An experiment that would settle the claim: keep the same occupancy decoder but replace the attention-based BEV construction with a depth-splatting (LSS-style) BEV builder on the OpenOcc benchmark, then compare the deformable-conv lifter against an MLP lifter. If the deformable-conv advantage shrinks to near zero, the paper's premise that BEV features contain recoverable vertical information is specific to attention-based BEV construction.","supporting_citations":[{"cited_title":"In: ICCV","cited_arxiv_id":null,"evidence_quote":"Supplies the BEVNet baseline, the 3D deformable attention alternative, and the OpenOcc benchmark."},{"cited_title":"In: ECCV","cited_arxiv_id":null,"evidence_quote":"Supplies the BEVFormer view transformer that produces the BEV features the lifter operates on."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"State-of-the-art occupancy baseline the method compares against on OpenOcc and SemanticKITTI."},{"cited_title":"In: ECCV (2020) 3, 8","cited_arxiv_id":null,"evidence_quote":"The LSS view transformation pipeline cited as the alternative way to build BEV features."}],"review_version":1}