{"id":"10b6a0e8-578f-4fe2-955c-97468d0f783b","arxiv_id":"2412.02075","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A Gaussian splatting pipeline reconstructs indoor scenes as separable objects and uses a trained completion model to fill in occluded surfaces zero-shot.","lead":"This paper presents a 3D reconstruction system that separates a scene into individual objects and fills in their hidden back sides using a neural network trained on large 3D object databases. It reports running about ten times faster than existing object-level reconstruction methods while producing watertight, separable object meshes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The zero-shot Object Surface Completion claim is not tested under a category holdout or on real partial scans; the occlusion/watertight claims therefore rest on an unvalidated generalization assumption.","rationale":"The reader's weakest assumption concerns the quality of instance masks and monocular priors. That is a real concern, but I see a more load-bearing issue specifically for the novelty claim: the Object Surface Completion module is the component that supposedly delivers watertight, separable meshes under occlusion, yet its zero-shot behavior is never tested against held-out object categories or real partial scans. The synthetic evaluation uses a small number of BlenderKit assets that may lie within the same broad distribution as the large ShapeNet/Objaverse training corpus, and the ShapeNet evaluation uses complete point clouds, which does not exercise the occlusion-completion capability at all. Without a category split or a real-world partial-scan test, the central claim that GOC completes unseen object surfaces in occluded scenarios is not empirically established. This is not an internal contradiction in the method, but it is a missing condition that must hold for the headline claim to be true. The paper itself flags ambiguity as a limitation, reinforcing that the generalization boundary is unknown. The concrete category-holdout test would settle the question: if OSC fails on unseen categories, the zero-shot characterization is invalid for those categories and the occlusion/watertight claim must be narrowed. Because the reader's verdict is already CONDITIONAL and my concern adds a further condition rather than overturning the paper entirely, I recommend keeping the verdict unchanged.","tokens_in":15813,"tokens_out":10055,"duration_ms":113402,"concrete_test":"Run a strict category-holdout evaluation of OSC: split the ShapeNet Core v2 and Objaverse training meshes into disjoint category sets, train OSC on a seen-category subset only, and evaluate on partial point clouds from held-out categories — including BlenderKit assets from classes not in the training set and, if possible, real ScanNet object crops produced by the GOC reconstruction pipeline. Report IoU, Chamfer Distance, and F-score separately for seen versus unseen categories. If unseen-category F-score or CD degrades by more than a pre-registered margin relative to seen categories, the zero-shot generalization claim in Contribution 3 is unsupported, and the watertight/occlusion claim must be scoped to training categories.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Contribution 1 and the Conclusion claim that GOC generates watertight, separable object meshes \"even in scenarios involving occlusion,\" and Contribution 3 claims a zero-shot OSC model with generalizability for unseen surface completion. The evidence for this is limited to (i) ShapeNet test-set reconstruction from complete point clouds (Sec. 6.4, Table 7) and (ii) five private synthetic BlenderKit scenes (Sec. 6.2, Tables 1-2). Neither test establishes zero-shot generalization to unseen object categories or to real, noisy, partially observed point clouds. OSC is trained on roughly 400k ShapeNet Core v2 and Objaverse meshes (Sec. 5.3), but the synthetic test objects are drawn from only about ten BlenderKit assets per scene, with no reported category split and no audit of category overlap with the training set. The ShapeNet test-set evaluation uses complete point clouds, so it does not test occlusion completion at all. Moreover, Table 1 shows that applying OSC makes object accuracy worse (0.0045 to 0.0062 for GOC w/o OSC versus GOC), consistent with the completion model replacing observed geometry with plausible prior shapes rather than preserving it; this makes the need for a held-out-category test more acute. The supplementary limitation statement (Sec. 5) concedes OSC \"may struggle with complex object models due to ambiguity,\" but no failure analysis or robustness numbers are provided. The central occlusion-related claim is therefore conditional on OSC generalizing to the one distribution where it is never evaluated: real reconstructed partial point clouds of objects outside the training categories.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Gaussian Object Carver (GOC), a 3D Gaussian Splatting (3DGS) framework for object-compositional scene reconstruction. GOC combines monocular normal/depth priors, multi-view photometric and geometric consistency, and an instance-segmentation feature field to reconstruct per-object geometry. A separate Object Surface Completion (OSC) model, trained on about 400k ShapeNet and Objaverse meshes, takes partial point clouds and outputs complete occupancy fields, producing watertight object meshes. The paper reports a more than 10x training-time speedup over SDF-based compositional baselines (ObjectSDF++, RICO) on synthetic scenes, competitive or better geometric accuracy, and a zero-shot completion module that requires no fine-tuning. Experiments include five private BlenderKit synthetic scenes, ScanNet real-world scenes, and a ShapeNet test-set evaluation of OSC.","tokens_in":16160,"tokens_out":3535,"duration_ms":35245,"significance":"If the reported results are robust, the framework would be a meaningful practical advance: it is, to my knowledge, the first 3DGS-based object-compositional reconstruction method, and the efficiency gain over SDF-based compositional methods is large and plausible given the rendering speed of 3DGS. The paper is also commendably transparent about implementation details, and the release of code (promised) and the extensive ablation over loss terms, OSC architecture, and masking strategies are strengths. However, the central novelty claim about zero-shot, occlusion-robust object completion is not yet supported by the experimental evidence: the ShapeNet evaluation uses complete point clouds, the synthetic scenes are a private five-scene benchmark with no category-overlap audit, and the real-world ScanNet evaluation explicitly excludes OSC. The quantitative support for OSC is further muddied by an inconsistency in Table 2. These gaps concern the load-bearing claims of the paper, so revision is required.","major_comments":[{"comment":"The zero-shot OSC generalizability claim (Contribution 3) is not validated under the conditions that the claim requires. The ShapeNet test-set evaluation in Supplementary Table 7 uses complete point clouds as input, so it tests surface reconstruction rather than completion from partial or occluded observations. The synthetic-scene experiments (Tables 1 and 2) are limited to about ten BlenderKit assets per scene, with no reported category split and no audit of category overlap with the ShapeNet/Objaverse training set. There is no held-out-category experiment and no test on real, noisy, partially observed point clouds. The supplementary Limitation section itself concedes the model 'may struggle with complex object models due to ambiguity,' which underscores that the generalization claim is currently an assumption. A held-out-category evaluation and/or a real partial-scan experiment is needed to support the zero-shot claim.","section":"§6.2, §6.4, Supplementary Table 7"},{"comment":"The text accompanying Table 2 contains a factual inconsistency that reverses the direction of the effect. The text states that 'our accuracy decreased slightly after applying completion (GOC w/o OSC), from 0.0073 to 0.0038,' but the table lists GOC w/o OSC Accuracy = 0.0038 and GOC Accuracy = 0.0073. Since Accuracy is a lower-is-better metric, this means applying OSC degraded accuracy by a factor of about two, not improved it. The same pattern is visible in Table 1 (0.0045 to 0.0062). These numbers should be corrected and the interpretation revised, because they directly affect whether OSC can be claimed to improve geometric fidelity or merely completeness.","section":"§6.2, Table 2"},{"comment":"The ScanNet experiment is missing a description of how instance-mask ground truth is obtained. Section 4.3 supervises the segmentation feature field with 'instance mask GT,' but Section 6.3 does not specify whether these masks come from the ScanNet annotations, a pretrained segmentation network, or manual labeling. The object-reconstruction metrics in Table 3 are directly conditioned on these masks, so without this information the real-world object-level results are not reproducible and their validity cannot be assessed.","section":"§6.3, §4.3"},{"comment":"The synthetic benchmark is private (five scenes with manually configured camera paths) and all reported numbers come from a single run with no error bars or multiple seeds. Since the headline efficiency claim (5% of the time of SDF-based methods) and the accuracy comparisons rest on this benchmark, the absence of variance information makes the quantitative comparisons fragile. Reporting per-scene results, multiple random seeds, or releasing the benchmark would materially strengthen the evidence; at minimum, the current single-run table does not support strong claims of superiority over RICO and ObjectSDF++.","section":"§6.2, Tables 1-2"}],"minor_comments":[{"comment":"The heading 'Segmentatin Regularization' contains a typo; it should read 'Segmentation Regularization.'","section":"§4.3"},{"comment":"The heading 'Obejct Surfaces Completion Model' contains a typo; it should read 'Object Surface Completion Model.'","section":"§5"},{"comment":"In Section 3, the phrase 'As shown in Fig. 7' appears, but the overview figure is Figure 2 in the main text; Figure 7 is in the supplementary material. The reference should be corrected.","section":"§3"},{"comment":"Equation (11) is ambiguous: the expression '0.5 · 1 − 0.5 × SDF(Qg)/Tiso' should be parenthesized as '0.5 * (1 - 0.5 * SDF(Qg) / Tiso)' to make the intended labeling-smoothing formula clear.","section":"§5.2, Eq. (11)"},{"comment":"In Table 1, the GOC and GOC w/ ShapeFormer rows have dashes ('-') in the scene-reconstruction columns, while the object-reconstruction columns are filled. This makes the table visually incomplete; the scene-reconstruction metrics for these variants should be reported, or the table should clarify that scene metrics are not applicable after object completion.","section":"§6.2, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core framework is promising. The main reason for major revision, rather than reject, is that the efficiency and reconstruction-quality contributions are plausible and the pipeline is novel; however, the zero-shot generalization claim is a central contribution and the current evidence is insufficient. The Table 2 inconsistency and the missing ScanNet mask details must be fixed. I would also encourage the authors to make the synthetic benchmark public, as the reliance on a private five-scene benchmark is a significant limitation for a paper whose main quantitative comparisons are on that benchmark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a system paper that combines object-compositional 3D Gaussian splatting with a learned surface completion transformer. The new bit is the integration and the camera-visibility masking for the OSC module; the components are borrowed from prior work, but I don't know another system doing exactly this in one pipeline. The headline efficiency claim is credible: 1h9m versus 18-21h for SDF-based baselines is a real order-of-magnitude win, and the qualitative results look genuinely better on the synthetic scenes.\n\nThe weakest part is the evidence for the zero-shot completion claim. OSC is trained on ShapeNet/Objaverse, then tested on five synthetic BlenderKit scenes and on the ShapeNet test set with complete point clouds. Neither test is a category holdout, and the BlenderKit assets have no reported category split or audit of overlap with the training set. The supplementary limitation admits the model 'may struggle with complex object models due to ambiguity,' but the abstract and conclusion still promise watertight meshes 'even in occluded areas.' On top of that, Table 1 shows accuracy gets worse after applying OSC on the full-observation set (0.0045 to 0.0062), which suggests the completion network can replace observed geometry with prior shapes. That tension needs a direct, quantitative discussion, not just a sentence in the appendix.\n\nThere is also a mechanical inconsistency in Table 2: the text says accuracy 'decreased from 0.0073 to 0.0038,' but the table shows 0.0038 for GOC w/o OSC and 0.0073 for GOC, meaning accuracy actually got worse. Easy fix, but it shakes confidence. And the ScanNet experiments don't specify how instance masks are obtained, which is a reproducibility gap.\n\nAll that said, the pipeline is coherent, the losses are standard regularizers, and the depth-ablation shows the geometry priors matter. The efficiency contribution is solid and the system is plausibly useful for real2sim and editing workflows. The gaps are experimental rather than conceptual: missing error bars, a private five-scene benchmark, and an overstated generalization claim.\n\nThis deserves a serious referee, not a desk reject. A solid revision would release the code and synthetic data, add multiple seeds, fix the table/text mismatch, and include a proper held-out-category completion evaluation on real partial scans. I'd send it out, with the generalization issue as the main request.","headline":"A useful integration of object-compositional Gaussian splatting with a learned completion module, but the zero-shot generalization claim is not actually tested; efficiency win is real.","tokens_in":16725,"tokens_out":1817,"would_cite":true,"duration_ms":19715,"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":"Gaussian Object Carver reconstructs indoor scenes as separable, watertight object meshes at over ten times the speed of SDF-based compositional methods, using a zero-shot completion module for occluded surfaces.","keywords":["3D Gaussian Splatting","object-compositional reconstruction","zero-shot surface completion","occupancy networks","monocular geometry priors","multi-view regularization","instance segmentation","watertight mesh extraction"],"falsifier":"Run GOC on a real ScanNet scene with masks produced by an automatic 2D segmenter and depth from a monocular estimator, then measure the fraction of completed object meshes that are actually watertight and separable; if small mask or depth errors cause object meshes to merge or leak, the claimed robustness to occlusion in real-world conditions fails.","tokens_in":15591,"feed_emoji":"🧊","tokens_out":8436,"duration_ms":69694,"temperature":0.7,"pith_summary":"Gaussian Object Carver (GOC) aims to show that object-compositional 3D scene reconstruction — splitting an indoor scene into individually editable objects — can be done with 3D Gaussian splatting instead of slow neural signed-distance fields. The paper's central claim is a tenfold speedup over existing compositional reconstructions while producing watertight, separable object meshes even when objects are partially occluded. The speed comes from replacing per-scene SDF optimization with differentiable Gaussian rasterization, and the watertightness comes from a separately trained zero-shot Object Surface Completion (OSC) module that fills in unobserved geometry from partial point clouds. If the claim holds, digital twins for simulation, AR/VR, and robotics could be built from real-world captures in about an hour per scene instead of a day, with objects that can be picked out, rearranged, and reinserted.","feed_headline":"Scene to watertight object meshes: Gaussian carver runs 10x faster","feed_subtitle":"Object-compositional reconstruction at one-twentieth the training cost, with zero-shot completion of occluded surfaces.","key_machinery":"The load-bearing object is the Object Surface Completion (OSC) model: a VAE-style occupancy network that encodes a partial surface point cloud (2048 farthest-point-sampled points, Fourier-position-encoded, through cross- and self-attention) into a latent set, and decodes per-query-point occupancy probabilities over a 3D grid; marching cubes then extracts a watertight mesh. In the reconstruction stage, the machinery is the alpha-blended depth and semantic-feature rendering of 3D Gaussians, with two regularization families — monocular priors (depth L1, alpha-weighted normal alignment) and multi-view geometry (photometric reprojection plus depth circular-consistency) — that keep the point cloud clean enough for the OSC module to complete.","core_discovery":"The paper introduces a two-stage pipeline. First, a 3D Gaussian representation of the scene is optimized with monocular depth and normal priors, photometric reprojection, and depth-consistency losses, while per-Gaussian semantic features are rendered and classified into instance masks, so the reconstructed geometry can be carved into objects. Second, the sampled point cloud of each object is fed into a zero-shot Object Surface Completion model, a point-cloud-conditioned occupancy network trained on roughly 400,000 watertight meshes, that predicts occupancy on a query grid and extracts a watertight mesh with marching cubes. The authors report that the full pipeline reconstructs a synthetic scene and completes all its objects in about 1 hour 9 minutes, roughly 5% of the time of the compared SDF-based methods, with better Chamfer distance and F-score on both object and scene metrics, and that the completion module transfers to unseen objects without fine-tuning.","pith_inferences":["A testable extension the authors do not run: replace their ground-truth instance-mask supervision with an automatic open-vocabulary 2D segmenter at inference time, to check whether object separation survives mask noise on real scans; their ScanNet comparison uses ground-truth instance masks, so this is the unresolved gap for fully automatic real-world use.","A second is to measure completion quality as a function of occlusion fraction; the paper's synthetic sparse setting fixes one viewing budget, but a sweep over viewing angles would show where the OSC prior starts to hallucinate rather than complete.","Because the OSC model encodes geometry only and deliberately ignores texture and appearance, a natural next step is to condition completion on multi-view appearance or language-image features — the authors themselves list this as future work — which would let the same module disambiguate symmetric or self-similar objects."],"forward_implications":["A full indoor scene can be turned into separable, editable object meshes in roughly an hour on a single consumer GPU, instead of tens of hours, making per-scene digital twins practical at scale.","Objects observed from a sparse set of viewpoints (about 30% of a full capture) still come out with higher completeness and F-score than dense SDF-based reconstruction, because the completion module fills occluded backsides.","The completion module, trained on a large mixed dataset of watertight meshes, generalizes zero-shot to reconstructed point clouds from Gaussian splatting, so no per-scene or per-category fine-tuning is needed.","Because objects are individually separated and watertight, downstream manipulation — deleting, duplicating, rearranging objects, or reinserting them into a simulator — becomes possible.","The monocular depth prior is the single most important regularizer: removing it cuts scene F-score from 0.8243 to 0.3483 in the real-world ablation."],"supporting_citations":[{"why":"Supplies the differentiable rasterizer and Gaussian primitive representation that the whole reconstruction stage builds on.","marker":"[17]"},{"why":"The main SDF-based object-compositional baseline whose accuracy and training time GOC is compared against.","marker":"[45]"},{"why":"The second SDF-based compositional baseline, used for the headline efficiency and accuracy comparison.","marker":"[20]"},{"why":"Source of the monocular depth and normal prior regularization idea for surface reconstruction.","marker":"[55]"},{"why":"Shows that depth and normal priors improve Gaussian splatting training and meshing, directly motivating GOC's monocular regularization.","marker":"[37]"},{"why":"Provides the geometry reprojection consistency loss used to enforce multi-view depth alignment.","marker":"[4]"},{"why":"Supplies the per-Gaussian learnable semantic feature rendering and mask supervision mechanism used for instance segmentation.","marker":"[53]"},{"why":"Its masking strategy is extended to 3D to train the OSC model from camera-visible partial point clouds.","marker":"[13]"},{"why":"Provides a large set of watertight object meshes used to train the completion module and to evaluate single-object surface reconstruction.","marker":"[3]"},{"why":"Supplies a large diverse set of object meshes that, combined with the other mesh dataset, gives the completion model zero-shot breadth.","marker":"[8]"}],"fun_headline_variants":["Gaussian Carver: 20x faster watertight scene meshing","Zero-shot completion turns Gaussians into watertight objects","Object-compositional Gaussian splatting with surface completion","GOC: object-level reconstruction at 5% training cost","Carve scenes into watertight objects with zero-shot completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline's headline results assume that every training view has accurate per-object instance masks and trustworthy monocular depth and normal priors; if those inputs are noisy, object separation erodes and the completion module cannot rescue the geometry.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian Carver: 20x faster watertight scene meshing","Zero-shot completion turns Gaussians into watertight objects","Object-compositional Gaussian splatting with surface completion","GOC: object-level reconstruction at 5% training cost","Carve scenes into watertight objects with zero-shot completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1366,"prompt_tokens":907,"completion_tokens":459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":375}},"tokens_in":523,"tokens_out":459,"duration_ms":4837,"temperature":1.0,"reasoning_tokens":375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:50:56.808351+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GOC on a real ScanNet scene with masks produced by an automatic 2D segmenter and depth from a monocular estimator, then measure the fraction of completed object meshes that are actually watertight and separable; if small mask or depth errors cause object meshes to merge or leak, the claimed robustness to occlusion in real-world conditions fails.","supporting_citations":[{"cited_title":"Objectsdf++: Improved object-compositional neural implicit surfaces","cited_arxiv_id":null,"evidence_quote":"The main SDF-based object-compositional baseline whose accuracy and training time GOC is compared against."},{"cited_title":"Rico: Regularizing the unobservable for indoor compositional reconstruction","cited_arxiv_id":null,"evidence_quote":"The second SDF-based compositional baseline, used for the headline efficiency and accuracy comparison."},{"cited_title":"Monosdf: Exploring monocu- lar geometric cues for neural implicit surface reconstruc- tion","cited_arxiv_id":null,"evidence_quote":"Source of the monocular depth and normal prior regularization idea for surface reconstruction."},{"cited_title":"Gaussian grouping: Segment and edit anything in 3d scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the per-Gaussian learnable semantic feature rendering and mask supervision mechanism used for instance segmentation."},{"cited_title":"Masked autoencoders are scalable vision learners","cited_arxiv_id":null,"evidence_quote":"Its masking strategy is extended to 3D to train the OSC model from camera-visible partial point clouds."},{"cited_title":"Objaverse: A universe of annotated 3d objects","cited_arxiv_id":null,"evidence_quote":"Supplies a large diverse set of object meshes that, combined with the other mesh dataset, gives the completion model zero-shot breadth."}],"review_version":1}