{"id":"c66b0a8d-24a5-4d19-b8a3-b7e16aa6e344","arxiv_id":"2508.13153","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"IGFuse jointly optimizes segmentation-aware Gaussian fields from multiple scans of rearranged scenes, producing complete, manipulable 3D reconstructions without inpainting.","lead":"IGFuse merges several scans of the same room taken after objects were moved, using the rearrangements to reveal hidden areas and build a complete 3D scene. The result is an editable Gaussian representation that renders new object arrangements without dense scanning or inpainting.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Known per-object rigid transforms and masks are assumed inputs; without them the claimed end-to-end fusion is unsupported.","rationale":"The reader's weakest assumption matches the main risk I see. The manuscript explicitly makes known rigid object transformations part of the method in Section 3.2 and per-object semantic identity part of the Gaussian state transfer in Section 3.3, yet it provides no estimator, initialization, or failure analysis for these inputs. Without them, the cross-scan alignment and the subsequent synthesis of novel object configurations cannot start. I find no deeper internal inconsistency: if the annotations are given, the joint optimization, bi-directional consistency, and co-pruning are plausible mechanisms for fusing Gaussian fields, and the fair-comparison appendix makes a reasonable attempt to address the multi-scan versus single-scan information gap. The unresolved issue is external validity, not mathematical soundness. A perturbation experiment on the input transforms and masks is the cheapest way to decide whether the method degrades gracefully or collapses without exact annotations. Because the paper is otherwise incomplete, with no code, no formal specification of the pseudo-intermediate state or co-pruning loss, and no error bars, I would keep the reader's CONDITIONAL verdict rather than move to accept or reject.","tokens_in":8292,"tokens_out":5480,"duration_ms":60811,"concrete_test":"Have the authors re-run the synthetic multi-scan protocol with per-object rigid transformations corrupted by realistic pose-estimation noise, e.g., independent N(0, 1 cm) translation and N(0, 1 deg) rotation per object, plus a 5-pixel erosion of the input segmentation masks, and report PSNR/SSIM for the same novel test state as Table 1. If PSNR drops by more than about 1 dB or falls below the single-scan baselines around 30 dB, then the known-annotation premise is load-bearing and the real-world claim must be scoped to settings that provide exact transforms and masks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 states, 'Using known rigid object transformations, we align the pair and fuse their information,' and Eq. (3) makes the synthesized novel state an explicit function of transformation T. Section 3.3's Gaussian State Transfer additionally requires each Gaussian to carry a semantic object identity o_i inherited from input segmentation masks. So the entire cross-scan fusion and novel-state synthesis pipeline is conditioned on two annotations that the paper never says how to obtain in real scenes: per-object rigid alignments and per-scan masks. The abstract and Figure 2 call the method end-to-end and claim real-world PSNR 27.18 'without complex pipelines,' but if the transforms are manually specified or tracked by external hardware, the comparison is not an end-to-end system and the real-data result does not transfer to unannotated captures. The text does not report a pose-estimation or mask-refinement module, an initialization, or a failure analysis under misalignment or mask error. The claimed multi-scan completion benefits are only realized after a geometric alignment that itself encodes the object rearrangement; if that alignment is given as input, a significant part of the claimed capability is assumed rather than learned. The central result, novel-state PSNR 36.93 on synthetic data and 27.18 on real data, measures the fused representation under oracle alignment, not the full claimed pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes IGFuse, a framework that reconstructs a complete and interactive 3D Gaussian scene by fusing several scans of the same scene captured under different object arrangements. Each scan is represented by a segmentation-aware Gaussian field, and scan pairs are aligned under assumed known rigid object transformations while enforcing bidirectional photometric and semantic consistency. A pseudo-intermediate Gaussian state and a collaborative co-pruning procedure are introduced to handle misalignment and remove inconsistent Gaussians. The authors evaluate novel-state synthesis on synthetic and real datasets, reporting PSNR 36.93 and 27.18 with large margins over single-scan baselines, and they provide additional segmentation-quality and object-completeness analyses.","tokens_in":8490,"tokens_out":5090,"duration_ms":52130,"significance":"If the claims hold under realistic input conditions, the paper would offer a useful alternative to inpainting-based or dense-capture pipelines for building manipulable scene representations: natural rearrangements between captures provide occlusion supervision, and the fused representation supports direct object-level transformation. Strengths include the clear multi-scan idea, bi-directional consistency losses, ablations isolating the pseudo-state contribution, and the fair-comparison experiments in Appendix H showing that merely training single-scan methods on additional scans does not close the gap. The main caveat is that the current manuscript measures performance conditional on known transforms and masks and does not specify or test the estimation of these quantities, so the headline 'end-to-end' capability is not yet established.","major_comments":[{"comment":"The entire fusion and novel-state synthesis pipeline assumes known rigid object transformations and per-scan segmentation masks as inputs: Eq. (5) makes each Gaussian's motion a function of the semantic label o_i and the transformation T^(o), and Section 3.2 states 'Using known rigid object transformations, we align the pair.' The manuscript does not describe how these quantities are obtained for real scans and provides no sensitivity analysis under misalignment or mask errors. Since the real-world PSNR 27.18 and the 'end-to-end' claim in the abstract and Figure 2 are predicated on this oracle input, the claimed capability is not demonstrated for unannotated captures. The authors should either supply a pose/mask estimation module or explicitly reframe the contribution as conditional on known annotations, and they should report robustness experiments.","section":"Section 3.2, Eqs. (2)-(5)"},{"comment":"The pseudo-intermediate Gaussian state is introduced as a central novelty, and Table 3 attributes a large PSNR improvement to it (35.55 to 36.93), but no definition, construction rule, optimization objective, or geometric relation to the two scan states is given anywhere in the supplied text. This makes the main contribution non-reproducible. Please provide a formal specification, including how pseudo-states are initialized and optimized and how the bidirectional losses are applied to them.","section":"Section 3.2, pseudo-intermediate state"},{"comment":"The joint objective L_joint is never expanded. The 'bi-directional photometric and semantic consistency' losses and the 'collaborative co-pruning' mechanism are described only verbally, without loss terms, thresholds, schedules, or an algorithm. As a result, the method cannot be reimplemented from the manuscript and the ablation in Table 3 cannot be reproduced. Add a complete method section with explicit equations for the photometric loss, the semantic feature loss, and the co-pruning criterion.","section":"Section 3.2, Eq. (2)"},{"comment":"No error bars, repeated-run statistics, or significance measures are reported for any quantitative result. The claimed gains are large, but with only a small number of scenes (the dataset and scene counts are not stated in the supplied text), the stability of PSNR 36.93/27.18 across random seeds, scan-pair choices, and initialization is unknown. Report mean and standard deviation over at least three runs and state the number of scenes and views.","section":"Tables 1-3 and 5-6"}],"minor_comments":[{"comment":"The section title 'Dual Guassian Convergence' contains a typo and should read 'Dual Gaussian Convergence'; the section describes discrepancies between fields but does not present quantitative convergence curves or a figure, so the reader cannot assess the claimed convergence behavior.","section":"Section 4.5"},{"comment":"The terms 'pseudo scene state' and 'pseudo-intermediate Gaussian state' are used interchangeably; please choose one definition and use it consistently throughout.","section":"Section 3.2"},{"comment":"The rendered segmentation feature S is not connected to the final mask prediction or to the semantic consistency loss; please explain how S is converted to masks and how the semantic loss is computed.","section":"Eq. (1)"},{"comment":"The text says 'incorporating multi-scans brings only slight improvements' for Table 6, but the table caption and surrounding discussion should clarify whether the baseline numbers are re-run under identical training schedules; otherwise the 'fair comparison' claim is hard to verify.","section":"Appendix H"}],"recommendation":"major_revision","confidential_remarks":"The submitted text appears to be an abbreviated version: the experimental setup section (datasets, metrics, baselines) and implementation details are missing, which makes the empirical contribution hard to evaluate. I would ask the editor to require the full version and, ideally, code or a detailed supplementary before a final decision. The input-assumption issue should also be pressed because it changes the scope of the claimed contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real contribution with a load-bearing assumption the paper under-sells. What's new is treating multiple scans of the same scene under natural object rearrangements as mutually completing observations, then jointly optimizing segmentation-aware Gaussian fields with bi-directional photometric and semantic consistency. The pseudo-intermediate Gaussian state and collaborative co-pruning are plausible mechanisms, and the ablation attributes most of the gain to the pseudo-state (35.10 to 36.93 PSNR). The appendix's fair comparison, giving single-scan baselines the same multi-scan training data, is the right check, and those baselines don't improve much. So the headline numbers (36.93 synthetic, 27.18 real) aren't just from feeding in more data.\n\nThe soft spot is exactly what the stress-test says. The method assumes known rigid per-object transformations and input segmentation masks; Eq. (3)-(5) make the synthesized state explicit functions of those annotations. The paper calls itself end-to-end and says it avoids complex pipelines, but it never says where these inputs come from in real scenes, reports no misalignment or mask-error sensitivity, and provides no estimator. If the real-data results are obtained with externally tracked or manually specified transforms, the comparison is against an oracle alignment, and the claim that this works on unannotated captures is unsupported. That doesn't kill the paper: fusion under known rearrangements is still a useful capability, e.g., for robotic simulators where object poses are tracked. But the authors need to scope the claim.\n\nSecondary issues: the pseudo-intermediate state and co-pruning are not formally specified; the ablation is only three rows; no error bars or repeated-run statistics; no code or data released. None of these undermines the central idea, but they make the large reported gains uncheckable from the text. The citation pattern is fine; the related work is relevant and not padded.\n\nBottom line: this deserves a serious referee. The novelty is real, and the core mechanism is coherent. A referee should push for code and data, a formal description of the pseudo-state, and sensitivity analysis on transformations and mask noise. If those come through, this could be a solid contribution; right now it is conditional, not broken.","headline":"IGFuse is a genuinely useful fusion idea under known object rearrangements, but the end-to-end claim outruns an input assumption that carries the method.","tokens_in":9053,"tokens_out":2435,"would_cite":true,"duration_ms":26559,"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":"IGFuse claims that fusing a few scans of a rearranged scene, through bi-directional photometric and semantic consistency with pseudo-intermediate alignment, reconstructs complete manipulable Gaussian scenes without inpainting or dense…","keywords":["3D Gaussian splatting","multi-scan fusion","interactive scene reconstruction","scene segmentation","novel state synthesis","object-level manipulation","real-to-sim transfer","novel view synthesis"],"falsifier":"Withhold the provided object poses and masks on the real-world split, estimate them by standard segmentation and registration, and measure novel-state PSNR; if it falls to the 21-22 dB single-scan baseline level, the claimed fusion depends on the privileged inputs rather than the method alone.","tokens_in":8068,"feed_emoji":"🖼️","tokens_out":6736,"duration_ms":61686,"temperature":0.7,"pith_summary":"This paper tries to establish that multi-scan fusion, rather than inpainting or per-object dense scanning, is the right route to complete interactive 3D scene reconstruction. It argues that when objects move between captures, the rearrangement reveals previously occluded surfaces, so each scan can supervise and complete the others. The method builds a segmentation-aware Gaussian field for each scan, aligns scan pairs through a pseudo-intermediate scene state, and enforces reciprocal photometric and semantic consistency while co-pruning inconsistent Gaussians. If the reported results hold, a few casually captured scans would suffice for rendering and object-level manipulation, which matters for robotics simulators and real-to-simulation transfer.","feed_headline":"Multi-scan fusion reconstructs interactive 3D scenes","feed_subtitle":"Jointly optimized Gaussian fields reach 36.93 PSNR on synthetic novel states, avoiding inpainting and dense capture.","key_machinery":"The load-bearing mechanism is pair-wise joint optimization of two segmentation-aware 3D Gaussian fields with reciprocal supervision: after alignment by known rigid object transformations, each scan's photometric and semantic renderings supervise the other, so surfaces occluded in one scan are completed by the other. A pseudo-intermediate scene state adds interpolation supervision that stabilizes alignment across configurations, and collaborative co-pruning removes Gaussians that remain inconsistent after alignment. Object-level manipulation is then produced by applying the known per-object rigid transformation to foreground Gaussians while leaving the background fixed.","core_discovery":"The authors claim that an interactive 3D Gaussian scene can be reconstructed end-to-end by jointly optimizing one segmentation-aware Gaussian field per input scan, under bi-directional photometric and semantic consistency, with a pseudo-intermediate scene state providing unified alignment and collaborative co-pruning refining geometry. On this basis, moving objects to novel configurations yields plausible consistent renderings without dense observations or multi-stage inpainting pipelines. The paper reports novel-state synthesis quality of 36.93 PSNR on synthetic data and 27.18 PSNR on real data, against 28-30 and 21-22 PSNR for single-scan baselines, and a 4-point mIoU gain in segmentation.","pith_inferences":["If the alignment assumption were relaxed by estimating object poses and masks jointly during fusion, the same reciprocal-supervision principle could extend to fully automatic capture.","The bi-directional consistency idea is portable to other scene representations, such as neural radiance fields or meshes, wherever multiple partial observations of a rearranged scene are available.","A direct extension would be measuring how novel-state PSNR degrades as the number of scans shrinks, which would reveal the minimum rearrangement needed for complete coverage."],"forward_implications":["Interaction-driven multi-scan capture can replace inpainting for completing occluded backgrounds and objects in Gaussian scenes.","Moving objects to arbitrary novel configurations renders plausibly and consistently, supporting object-level scene editing without dense per-object capture.","Cross-scan mutual supervision improves segmentation accuracy at object boundaries, reported as 91.0 mIoU versus 86.8 for a single-scan baseline.","The few-scan, no-inpainting setting lowers the capture cost for real-to-simulation transfer used in robotics training.","Simply adding more scan data to single-scan baselines yields little gain; the consistency and alignment mechanism, not the extra data alone, drives the improvement."],"supporting_citations":[{"why":"introduces 3D Gaussian Splatting, the underlying scene representation each per-scan field is built from.","marker":"Kerbl et al. 2023"},{"why":"defines segmented Gaussian Splatting with per-Gaussian features and alpha-blending, the formulation IGFuse inherits and the main baseline it improves.","marker":"Ye et al. 2023"},{"why":"DecoupledGaussian, the representative single-scan inpainting-based interactive reconstruction baseline it compares against.","marker":"Wang et al. 2025"},{"why":"GaussianEditor, a baseline segmentation/editing approach that projects 2D masks onto Gaussians.","marker":"Chen et al. 2024"},{"why":"GaussianCut, a graph-cut segmentation baseline used in the quantitative comparison.","marker":"Jain, Mirzaei, and Gilitschenski 2024"},{"why":"RoboGSim, an interactive Gaussian simulator whose dense per-object capture requirement the paper targets as a limitation.","marker":"Li et al. 2024b"},{"why":"SplatSim, another Gaussian real-to-sim approach that relies on dense capture, motivating the lightweight multi-scan alternative.","marker":"Qureshi et al. 2024"}],"fun_headline_variants":["IGFuse fuses scans for interactive 3D scene synthesis","Multi-scan Gaussian fusion enables novel 3D scene states","End-to-end scan fusion beats inpainting for interactive 3D","Object-level 3D control via multi-scan Gaussian fusion","Fusing scans reconstructs 3D scenes with occlusion handling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole reconstruction assumes every scan already comes with correct per-object rigid transformations and segmentation masks, and no image-based estimator for these inputs is supplied in the paper.","fun_headline_variants_meta":{"raw":{"variants":["IGFuse fuses scans for interactive 3D scene synthesis","Multi-scan Gaussian fusion enables novel 3D scene states","End-to-end scan fusion beats inpainting for interactive 3D","Object-level 3D control via multi-scan Gaussian fusion","Fusing scans reconstructs 3D scenes with occlusion handling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1332,"prompt_tokens":889,"completion_tokens":443,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":354}},"tokens_in":505,"tokens_out":443,"duration_ms":4579,"temperature":1.0,"reasoning_tokens":354,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:14:40.587637+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Withhold the provided object poses and masks on the real-world split, estimate them by standard segmentation and registration, and measure novel-state PSNR; if it falls to the 21-22 dB single-scan baseline level, the claimed fusion depends on the privileged inputs rather than the method alone.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GaussianEditor, a baseline segmentation/editing approach that projects 2D masks onto Gaussians."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GaussianCut, a graph-cut segmentation baseline used in the quantitative comparison."}],"review_version":2}