{"id":"13698e3f-e135-4f26-ab8d-41791ed7aa3e","arxiv_id":"2507.18331","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"SGCDet combines per-view 3D deformable attention with box-supervised sparse volume refinement to achieve state-of-the-art multi-view indoor 3D object detection without ground-truth scene geometry.","lead":"This paper introduces SGCDet, a neural network that detects 3D objects in indoor scenes from multiple ordinary photos without needing a depth sensor or a 3D scene model. It reports large accuracy gains over prior image-only methods while using less memory and training time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is unverified because GO-N3RDet, a recently cited baseline that appears not to use ground-truth geometry, is absent from all comparisons.","rationale":"The paper's headline contribution is an empirical SOTA claim, and that claim depends on comparing against every relevant method that does not use ground-truth geometry. The paper mentions GO-N3RDet and NeRF-Det++ in related work, so they are known to the authors, but they are excluded from the quantitative comparisons without explanation. GO-N3RDet in particular is a CVPR 2025 method that enhances NeRF-Det through geometric optimization; nothing in its title or the paper's description indicates it uses ground-truth geometry supervision. If it is a no-geometry method, its omission is a direct threat to the central claim, not a peripheral reproducibility issue. The margin over MVSDet is large, but the missing baseline could be even stronger. The reader's own rationale flagged this missing-baseline issue, though their formal 'weakest assumption' focused on pseudo-labels. I consider the missing baseline more load-bearing because the pseudo-label concern is at least partially addressed by the noisy-label ablation, while the baseline omission is unaddressed. The appropriate verdict remains CONDITIONAL: the paper's design and ablations are coherent, but the SOTA claim cannot be accepted until the missing comparisons are supplied. Hence the verdict is unchanged from the reader's CONDITIONAL assessment.","tokens_in":13671,"tokens_out":9971,"duration_ms":103430,"concrete_test":"Check GO-N3RDet's training protocol: if it does not supervise with ground-truth depth or geometry, re-evaluate it on the ScanNet validation set under the same protocol used for SGCDet (40 training images, 100 test images, ResNet-50 backbone, mAP@0.25 and mAP@0.5), and add it to Table 1. Also verify whether NeRF-Det++ uses ground-truth depth supervision; if it does not, evaluate it as well. If either method reaches or exceeds 61.2 mAP@0.25, the SOTA claim is falsified. If both fall below, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is state-of-the-art among methods that do not use ground-truth scene geometry for supervision (Abstract, §4.3). The related work explicitly cites GO-N3RDet [17] and NeRF-Det++ [9] as NeRF-Det enhancements, yet neither appears in Tables 1–3. If GO-N3RDet's 'geometry optimized' NeRF-based pipeline is trained without ground-truth depth or geometry—which the paper's own description suggests—then omitting it means the SOTA claim has not been tested against a directly relevant recent baseline. The reported margin over MVSDet is 5.0 mAP@0.25, so a missing strong baseline could plausibly fall within or beyond that margin. The reader's identified weakest assumption, occupancy pseudo-label sufficiency, is partially mitigated by Table 7's 15% label-drop robustness experiment, which shows only a 0.5 mAP@0.25 drop under incomplete labels. The missing baseline comparison is therefore more load-bearing for the headline claim than the pseudo-label concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SGCDet, a multi-view indoor 3D object detector that combines a geometry- and context-aware aggregation module (3D deformable attention plus multi-view attention) with a coarse-to-fine sparse volume construction that selects top-k voxels by predicted occupancy. Occupancy supervision is derived from ground-truth 3D bounding boxes, avoiding explicit scene-geometry supervision. On ScanNet, ScanNet200, and ARKitScenes, SGCDet reports higher mAP than previous geometry-supervision-free methods, particularly MVSDet (e.g., 61.2 vs 56.2 mAP@0.25 on ScanNet), while lowering training/inference memory and time. Ablations in Tables 4–8 support the main design choices.","tokens_in":13952,"tokens_out":8509,"duration_ms":82111,"significance":"If the reported results hold, SGCDet would be a practical advance for image-based indoor detection: it removes the need for ground-truth geometry, improves accuracy over MVSDet by 5.0 mAP@0.25, and cuts memory/time by roughly half. The ablations are systematic and the public code promise supports reproducibility. However, the headline SOTA claim is currently under-supported because two recently cited baselines (GO-N3RDet, NeRF-Det++) are absent from all comparisons, and the absence of variance estimates makes some smaller ablative differences hard to interpret.","major_comments":[{"comment":"The state-of-the-art claim is not verified against the full set of relevant baselines. Section 2 cites GO-N3RDet [17] and NeRF-Det++ [9] as recent methods that 'further enhance NeRF-Det through semantic and geometric constraints,' and nothing in the descriptions indicates they use ground-truth geometry supervision. Yet neither is included in the quantitative comparisons in Tables 1–3. Because the paper's central claim (Abstract, §4.3) is superiority over approaches that do not rely on ground-truth geometry, omitting these directly relevant baselines leaves the claim unproven; the 5.0 mAP@0.25 margin over MVSDet does not rule out a missing baseline being competitive. Please add comparisons to these methods, or explicitly demonstrate that they fall outside the 'no ground-truth geometry' class, and adjust the SOTA claim accordingly.","section":"§2, §4.3, Tables 1–3"},{"comment":"All experiments are single runs without variance estimates or significance tests. The main SOTA gain (5.0 mAP@0.25 over MVSDet, Table 1) is large, but several design choices are based on much smaller differences: Table 4 (a) vs (b) is 0.2 mAP, and Table 5 (a) vs (e) is 0.2 mAP. Without repeated seeds or error bars, it is unclear whether these differences are meaningful or within run-to-run noise. Please report at least three seeds (mean±std) for the key comparisons, or otherwise justify the stability of the reported numbers.","section":"§4.3–4.5, Tables 1–8"},{"comment":"The efficiency comparison is not on an equal footing. The footnote to Table 1 states that * results are directly cited from [30,40], while the text says 'The computational cost is measured on a single NVIDIA A6000 GPU.' If the baseline memory/time/FPS numbers are taken from other papers, they may reflect different hardware, software, or batch settings, so the claimed reductions (42.9% training memory, 47.2% training time, 50% inference memory, 40.8% faster inference) are not established. Please measure all baselines under identical conditions, or clearly separate reproduced numbers and add appropriate caveats.","section":"Table 1, §4.3"},{"comment":"The occupancy pseudo-labels come from ground-truth bounding boxes and therefore cover only annotated object interiors. The top-25% voxel selection could in principle discard voxels belonging to unannotated objects or retain free space. Table 7's 15% label-dropout experiment is a useful robustness check, but it does not directly validate the claim in Section 4.5 that 'the top 25% selection for refinement ensures sufficient coverage of occupied regions, including areas not annotated by pseudo-labels.' Please provide direct evidence (e.g., recall of occupied voxels as a function of selection ratio, or a test with a large fraction of unannotated objects) or soften the claim.","section":"§3.2, §4.5, Table 7"}],"minor_comments":[{"comment":"The phrase 'state-of-the-art performance' overstates what is demonstrated, since on ScanNet200 (Table 2) only two baselines are compared; please qualify the claim to the evaluated methods.","section":"Abstract, §5"},{"comment":"In Eq. (2), the notation 'top k%' uses k as both a percentage and a count; please adopt a single notation (e.g., alpha for the fraction).","section":"§3.2"},{"comment":"The sentence 'The models are trained on NVIDIA A6000 GPUs' should specify the number of GPUs, per-GPU batch size, and total batch size, since training time is reported as a contribution.","section":"§4.2"},{"comment":"The original and reproduced baseline rows are both listed; this is useful but potentially confusing. Please label them clearly, e.g., 'original' vs 'reproduced with origin recentering'.","section":"Table 3"},{"comment":"Reference [7] lists the first author as 'A Waswani'; the correct spelling is 'A. Vaswani'.","section":"References"},{"comment":"The caption explains that the depth dimension is omitted for clarity; it would help to also mark the image boundaries in each view.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The missing baseline comparison is the most serious issue: GO-N3RDet and NeRF-Det++ are cited in Section 2 as recent enhancements of NeRF-Det, and their omission from all quantitative tables directly weakens the SOTA claim. The authors should either add these comparisons or provide a clear justification for exclusion. The lack of variance estimation, especially for small ablative differences, is a recurring problem that should be addressed editorially. There is no indication of misconduct; the omissions need to be justified in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid systems contribution to multi-view indoor 3D detection from images. The core idea — replace fixed single-point sampling in voxel lifting with per-view 3D deformable attention plus a lightweight multi-view attention — is well motivated and the ablations show it is doing real work. The second piece, selecting the top 25% occupied voxels for refinement using a box-derived occupancy pseudo-label, is also well engineered and delivers the claimed ~50% training/inference savings with only a small accuracy cost. The headline numbers on ScanNet (61.2 mAP@0.25, 35.2 mAP@0.5) are good and the gains over MVSDet are large.\n\nMy main pushback is on the SOTA claim. The paper explicitly cites GO-N3RDet and NeRF-Det++ as recent NeRF-Det enhancements, but neither appears in any comparison table. If those methods operate without ground-truth geometry — which the paper's own text seems to suggest — then the headline \"state-of-the-art among methods without GT geometry\" is not actually tested. That is a load-bearing omission, not a nit. The margin over MVSDet is 5 mAP points, so a missing strong baseline could plausibly close or reverse the gap. Referees should ask for a fair comparison, or a clear statement of why those methods are excluded.\n\nOther soft spots are secondary. All reported numbers are single runs; no variance or significance testing. That is common in this subfield but worth noting given the claims. The ARKitScenes table mixes original numbers with re-implemented baselines; the re-implementations are explained, but the discrepancy between original and reproduced numbers is large enough that someone should double-check the protocol. The pseudo-label occupancy supervision is properly auxiliary and not circular, and the robustness test with 15% dropped/scaled boxes is reassuring. The concern that unannotated objects might be missed by the top-k selection is mitigated by that same test, though not fully eliminated.\n\nThe code is promised, the math in Eqs. 3–6 is clean, and the ablation tables are among the most thorough I've seen in this line of work. I'd send it to peer review with a request that the authors add the missing baseline comparisons and report variance if they can. It is a real advance on the engineering side and the paper as a whole is worth careful reading.","headline":"Solid architecture for multi-view indoor 3D detection with large reported gains, but the SOTA claim is undercut by omitting recently cited baselines like GO-N3RDet from the comparisons.","tokens_in":14462,"tokens_out":2632,"would_cite":true,"duration_ms":24410,"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":"SGCDet claims that multi-view indoor 3D object detection becomes both more accurate and cheaper when the 3D volume is built adaptively: deformable feature aggregation in a lifted 3D space, refinement of only the most occupancy-likely…","keywords":["multi-view 3D object detection","indoor scene understanding","adaptive 3D volume construction","sparse volume refinement","occupancy prediction","deformable attention","bounding-box pseudo-labels","ScanNet"],"falsifier":"On a validation scene, compute the occupancy ranking from the trained occupancy head and compare the top-25% voxels against a dense occupancy map built from all visible objects, including unannotated ones; if mAP under 25% selection falls well below mAP under 100% selection in scenes with heavy unannotated clutter, the box-derived pseudo-labels are not sufficient to rank object voxels.","tokens_in":13504,"feed_emoji":"📦","tokens_out":5882,"duration_ms":55562,"temperature":0.7,"pith_summary":"SGCDet is a multi-view indoor 3D object detector that tries to show that a 3D volume built from posed images can be both more accurate and far cheaper when the network is allowed to choose where to look. Its central proposal is adaptive volume construction: instead of sampling image features at fixed projected points and refining every voxel, the network aggregates features inside deformable 3D regions and refines only voxels judged likely to hold objects. The occupancy judgment is supervised by pseudo-labels derived from 3D bounding boxes, so the whole pipeline needs image labels and box labels rather than ground-truth scene geometry. On ScanNet the method reports mAP@0.25 of 61.2 and mAP@0.50 of 35.2, outperforming the previous geometry-free state of the art MVSDet by 5.0 and 3.9 points while also lowering training memory, training time, inference memory, and inference time.","feed_headline":"Box labels alone power top indoor 3D detection at 61.2 mAP","feed_subtitle":"SGCDet refines only likely-object voxels, cutting training memory 43% and inference memory 50% while beating prior image-only detectors.","key_machinery":"The load-bearing mechanism is a coarse-to-fine sparse volume construction teamed with a geometry and context aware aggregation module. At each refinement stage an occupancy head scores voxels of an upsampled coarse volume, the network keeps the top-k% voxels for feature refinement, and the aggregation module lifts image features by 3D deformable attention: each voxel is projected into a lifted 3D pixel space where its sampled feature acts as a query and deformable offsets gather context, followed by multi-view attention that re-weights the views. The occupancy head is trained with pseudo-labels obtained by filling ground-truth 3D bounding boxes, so no depth map or reconstructed mesh is needed.","core_discovery":"SGCDet claims that the two usual costs of voxel-based multi-view detection — fixed receptive fields and dense computation over empty space — can be removed together by making volume construction adaptive in both feature space and voxel space. The geometry and context aware aggregation treats each voxel's projected feature as a query in a 3D deformable attention over lifted image features, letting the network pull information from a flexible neighborhood in each view and then weight views by attention; the sparse volume construction builds a coarse volume, predicts per-voxel occupancy, and refines only the top 25% of voxels at each stage, with occupancy supervised by binary labels that mark voxels inside any ground-truth 3D box. The paper reports that this combination achieves state-of-the-art results among methods trained without ground-truth geometry on ScanNet, ScanNet200, and ARKitScenes, and that on ScanNet it exceeds MVSDet by 5.0 mAP@0.25 and 3.9 mAP@0.50 while reducing training memory, training time, inference memory, and inference time by 42.9%, 47.2%, 50%, and 40.8%.","pith_inferences":["An inference: if box-derived occupancy pseudo-labels stay reliable in scenes with many unannotated objects, the same coarse-to-fine sparse selection could be dropped into other voxel-based perception tasks such as semantic occupancy prediction or scene completion that today rely on full geometry supervision; the paper does not test that transfer.","An inference: the top-25% selection creates a hidden coupling between annotation density and accuracy, because adding more annotated boxes changes the occupancy supervision and therefore which voxels get refined; a dataset with sparse labels could silently degrade the sparse volume before the detection head sees it.","An inference: because the deformable attention operates in a lifted 3D pixel space rather than on the 2D image, the same module could serve monocular or few-view settings where depth estimates are noisy, as long as the depth distribution carries enough signal; the paper only evaluates the full multi-view pipeline."],"forward_implications":["Because supervision needs only 3D bounding boxes, the method can be applied to datasets that have image poses and box labels but no depth or reconstructed geometry.","The voxel-selection ratio is a direct accuracy-efficiency dial; the ablations show that 10% selection loses accuracy while 25% preserves it, so downstream systems can trade speed for recall.","The reported cost reductions versus MVSDet mean the accuracy gain is not bought with compute: training memory falls 42.9%, training time 47.2%, inference memory 50%, and inference time 40.8%.","On ScanNet200's 200-category split and ARKitScenes' oriented boxes, the method generalizes beyond the 18-class ScanNet setting; the higher-resolution SGCDet-L reaches an 80x80x32 volume and outperforms the geometry-supervised CN-RMA on ARKitScenes."],"supporting_citations":[{"why":"Establishes the image-to-voxel projection baseline and the anchor-free detection head that SGCDet builds on.","marker":"[29]"},{"why":"The previous state of the art without ground-truth geometry; its plane-sweep depth module is the main accuracy and efficiency baseline.","marker":"[40]"},{"why":"Defines the geometry-supervised baseline and the ScanNet200/ARKitScenes evaluation protocol, including coordinate re-centering.","marker":"[31]"},{"why":"A geometry-free baseline that uses opacity post-processing, which SGCDet replaces with learned sparse refinement.","marker":"[38]"},{"why":"A geometry-supervised, multi-stage baseline that SGCDet outperforms despite not using ground-truth geometry.","marker":"[30]"},{"why":"Supplies the 3D deformable attention mechanism that the geometry and context aware aggregation adapts for indoor scenes.","marker":"[12]"},{"why":"The ScanNet dataset and train/test split used for the main detection results.","marker":"[6]"},{"why":"The ScanNet200 dataset used for the 200-category generalization test.","marker":"[27]"},{"why":"The ARKitScenes dataset used for oriented 3D bounding box evaluation.","marker":"[2]"}],"fun_headline_variants":["Adaptive sparse volumes: 50% less memory, 5 mAP gain","Indoor 3D detection from box labels alone via adaptive volumes","Sparse volume construction: efficient multi-view 3D detection","Adaptive voxel refinement: beat MVSDet with 50% less memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that filling ground-truth 3D bounding boxes marks the voxels worth refining, so if a scene contains many unannotated objects or clutter outside boxes, the top-25% occupancy selection could discard object voxels or keep empty ones.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive sparse volumes: 50% less memory, 5 mAP gain","Indoor 3D detection from box labels alone via adaptive volumes","Sparse volume construction: efficient multi-view 3D detection","Adaptive voxel refinement: beat MVSDet with 50% less memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001139,"raw_usage":{"total_tokens":4745,"prompt_tokens":981,"completion_tokens":3764,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":3684}},"tokens_in":597,"tokens_out":3764,"duration_ms":30395,"temperature":1.0,"reasoning_tokens":3684,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:14:04.504363+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a validation scene, compute the occupancy ranking from the trained occupancy head and compare the top-25% voxels against a dense occupancy map built from all visible objects, including unannotated ones; if mAP under 25% selection falls well below mAP under 100% selection in scenes with heavy unannotated clutter, the box-derived pseudo-labels are not sufficient to rank object voxels.","supporting_citations":[{"cited_title":"ImV oxelNet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection","cited_arxiv_id":null,"evidence_quote":"Establishes the image-to-voxel projection baseline and the anchor-free detection head that SGCDet builds on."},{"cited_title":"MVSDet: Multi- view indoor 3d object detection via efficient plane sweeps","cited_arxiv_id":null,"evidence_quote":"The previous state of the art without ground-truth geometry; its plane-sweep depth module is the main accuracy and efficiency baseline."},{"cited_title":"Im- GeoNet: Image-induced geometry-aware voxel representa- tion for multi-view 3d object detection","cited_arxiv_id":null,"evidence_quote":"Defines the geometry-supervised baseline and the ScanNet200/ARKitScenes evaluation protocol, including coordinate re-centering."},{"cited_title":"NeRF-Det: Learning geometry-aware volumetric representation for multi-view 3d object detection","cited_arxiv_id":null,"evidence_quote":"A geometry-free baseline that uses opacity post-processing, which SGCDet replaces with learned sparse refinement."},{"cited_title":"CN-RMA: Combined network with ray marching aggrega- tion for 3d indoor object detection from multi-view images","cited_arxiv_id":null,"evidence_quote":"A geometry-supervised, multi-stage baseline that SGCDet outperforms despite not using ground-truth geometry."},{"cited_title":"DFA3D: 3d deformable attention for 2d-to-3d feature lifting","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D deformable attention mechanism that the geometry and context aware aggregation adapts for indoor scenes."},{"cited_title":"ScanNet: Richly-annotated 3d reconstructions of indoor scenes","cited_arxiv_id":null,"evidence_quote":"The ScanNet dataset and train/test split used for the main detection results."}],"review_version":2}