{"id":"2480f522-8e83-4e9d-9f84-221b83330cd4","arxiv_id":"2608.01492","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A training-free graph-cut method selects 3D objects from Gaussian splatting scenes using sparse user scribbles, reaching 92.2 mIoU on NVOS with three interaction views.","lead":"GaussianSelector picks out a 3D object from a reconstructed scene using just a few user scribbles on one to three views, without retraining or a segmentation network. It builds a graph over Gaussian splats and cuts it to propagate the scribbles to a full 3D selection, hitting quality close to heavier multi-view SAM-based methods at a fraction of the compute.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's iterative foreground-restriction loop is undocumented and unablated; the reported mIoU gains may stem from this step rather than the proposed graph-cut formulation.","rationale":"I read the paper in good faith. The central claim is that a training-free, scribble-driven graph-cut on superpoints reaches 85.3/92.2 mIoU. The method description in Sec. 3.5 emphasizes a single graph-cut optimization. Algorithm 1, however, includes an iterative foreground-restriction loop that is not described in the text or accounted for in the ablation. This is a concrete internal inconsistency and a reproducibility risk: if the loop is essential, the paper's stated contribution (the energy formulation) is incomplete; if it is not essential, it should be removed or clarified. The reader's weakest assumption about seed lifting (Eq. 6) is reasonable, but I think the undocumented loop is more load-bearing because it affects every result and cannot be inspected. Therefore, I partially agree with the reader: both point to missing algorithmic details, but the loop is the more critical omission. The verdict remains CONDITIONAL pending an ablation and specification.","tokens_in":11757,"tokens_out":7262,"duration_ms":80309,"concrete_test":"Implement GaussianSelector on the NVOS benchmark exactly as in Algorithm 1 but with the repeat loop disabled (i.e., perform a single graph cut per interaction round). Use the same hyperparameters and initial scribbles as the paper. Compare the mIoU with the full Algorithm 1. If the single-shot variant drops by more than ~1 point (or if the loop is required to reach 85.3), the iterative foreground-restriction must be reported and ablated; if it makes no difference, the concern is resolved. Additionally, report the loop's iteration count and convergence criterion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core contribution is the graph-cut energy in Eq. (1), and Sec. 3.5 describes the method as a single optimization: 'Once the optimal labeling L* is obtained, it is broadcasted back...' However, Algorithm 1 (lines 8-13) contains a repeat loop that, after each min-cut, rebuilds the superpoint graph on the foreground subgraph {k: L*_k = F} and re-estimates the likelihood models 'on the restricted graph.' This loop is never mentioned in the method text, its convergence criterion is unspecified, and it is not ablated in Table 4. If this loop modifies the selection beyond what the single graph cut achieves, then the claimed performance (85.3 mIoU single-view, 92.2 mIoU with 3 views) is not produced by the described energy minimization. The loop is also ambiguous: restricting to the foreground nodes would leave no background seeds to re-estimate p_B, so the intended operation is unclear. Without a precise description or ablation, the central claim that sparse scribbles suffice to match dense multi-view SAM-lifting is not reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"GaussianSelector proposes a training-free interactive 3D object selection method for 3D Gaussian Splatting. From sparse user scribbles on one or a few views, it builds a scene-native superpoint graph: Gaussians are reparameterized with a Canonical Axis Color (CAC) descriptor, coarsened via Leiden clustering, and connected with continuity-weighted k-NN edges. Scribbles are lifted to Gaussian-level seeds through transmittance-weighted coverage, aggregated to superpoint seeds, and used to fit GMM appearance models. Selection is posed as a binary submodular graph-cut energy (Eq. 1) solved exactly by s-t min-cut. The paper reports 85.3 mIoU on NVOS with a single scribble view and 92.2 mIoU with three interaction rounds, competitive with SAM-lifting baselines that use all views, while requiring no retraining and substantially less computation.","tokens_in":12059,"tokens_out":5245,"duration_ms":49599,"significance":"If the central claim holds, the contribution is significant: a training-free, GPU-light, sparse-scribble method that matches dense multi-view SAM-based selection would lower the barrier for interactive 3DGS editing and asset extraction. Strengths include the clean separation of scribble-independent scene encoding from interactive inference, the use of an exact graph-cut solver, and concrete runtime measurements showing large gains over feature-field and SAM-lifting baselines. The method is falsifiable on standard benchmarks and the component ablation in Table 4, once clarified, can support the design. However, the current manuscript leaves a load-bearing algorithmic loop unspecified and the multi-round protocol under-controlled, so the headline numbers are not yet reproducible as described.","major_comments":[{"comment":"Algorithm 1 lines 8–13 contain a repeat loop that is never described in the method text. Section 3.5 states: 'Once the optimal labeling L* is obtained, it is broadcasted back...', presenting Eq. (1) as a single optimization. In contrast, Algorithm 1, after each min-cut, restricts the superpoint graph to the foreground subgraph {k: L*_k = F} and re-estimates the likelihood models on the restricted graph. The convergence criterion is unspecified, the number of iterations is not reported, and this loop is not ablated in Table 4. Moreover, restricting to foreground nodes removes all background seeds, making re-estimation of p_B undefined unless additional context nodes are retained, which is not described. As written, the reported 85.3/92.2 mIoU may be produced by an undocumented iterative post-processing step rather than by the energy in Eq. (1). Please either remove the loop, specify it pr","section":"Algorithm 1 / §3.5"},{"comment":"The multi-round refinement protocol is not reproducible from the manuscript. Section 4.2 states that 'users are allowed to select informative views and inject scribbles', but no procedure is given for selecting these views, how many strokes are used, how ambiguous cases are resolved, or how the '2 Rounds' and '3 Rounds' rows were obtained. Since mIoU is reported without standard deviations or multiple annotators, the 89.6 and 92.2 numbers could depend strongly on favorable view selection. The paper should specify a fixed view schedule, report per-task breakdowns, and ideally include multiple-user variability. Without this, the central 'fewer views' claim is not yet a controlled comparison.","section":"Table 1 / §4.2"},{"comment":"The comparison to GaussianCut under sparse views is qualitative only. Table 1 reports GaussianCut only with ALL views, while the text and Fig. 3 claim GaussianCut degrades when given fewer views. Since GaussianCut is the most directly comparable graph-cut baseline on 3DGS, the paper should report quantitative IoU for GaussianCut under the same 1-view and 3-view settings (or explain why such a comparison is not possible). Without this, the claim that GaussianSelector matches dense multi-view methods using fewer views is not directly established for the closest baseline.","section":"Table 1 / Fig. 3"}],"minor_comments":[{"comment":"The check marks in Table 4 appear inconsistent with the text in §4.4. The text says adding graph propagation improves from 33.2 to 61.0, and adding unary modeling reaches 79.6, which implies the 'w/o unary' row should have Unary=×, Graph=✓, and the 'w/o graph' row should have Unary=✓, Graph=×. The table as printed shows the opposite. Please fix the row labels or the marks.","section":"Table 4"},{"comment":"The affine normalization parameters m and s are described only in words ('midpoint between seed medians' and 'adaptive scale determined by separation') and the target seed confidence of 0.95 is listed in Implementation Details but not connected to Eq. (9). Please give the exact formula for s and how the 0.95 confidence is used.","section":"§3.4, Eq. (9)"},{"comment":"The distances d_x, d_c, d_o in Eq. (4) are called 'normalized', but no normalization procedure is defined. Also, the gating threshold is a quantile of which distribution? Specify these details for reproducibility.","section":"§3.2"},{"comment":"Equation (6) uses α_ip and T_ip without defining the pixel index convention and how the per-view threshold is set. The phrase 'per-view threshold' is not specified numerically or algorithmically; please clarify how thresholds are chosen and how the majority vote tie rule works.","section":"§3.3"},{"comment":"All experiments are reported without error bars or significance tests. At minimum, report standard deviations over multiple runs or multiple view selections for the key Table 1 rows.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript does not mention code release. Given that the central claims hinge on an under-specified algorithm loop and a user-driven evaluation protocol, code or a detailed public protocol would be important for reproducibility. The framing also should more carefully distinguish the novelty from GaussianCut; the superpoint coarsening and CAC descriptor are the substantive differences, and the paper would be strengthened by a controlled sparse-view comparison with that baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GaussianSelector is a genuine attempt to make interactive object selection in 3DGS cheap and sparse: Leiden superpoints, a CAC appearance feature, transmittance-weighted scribble lifting, and a graph-cut energy. The combination is new, the energy is submodular and exactly solvable, and the runtime numbers (0.2 min per query, CPU-only) are the real selling point. If the method works as described, it is a useful tool for scene editing and asset extraction, even if the quality is only on par with GaussianCut rather than better.\n\nThe paper does a few things well. The ablation separates unary vs graph contributions, the qualitative figures show sensible behavior, and the user study, while small, supports the \"fast enough to iterate\" claim. The writing is clear about the single-view and multi-round settings. The authors also correctly position the result as competitive, not state-of-the-art, on NVOS.\n\nThe biggest soft spot is Algorithm 1's repeat loop (lines 8-13). The method section describes one s-t min-cut, but the pseudocode iteratively rebuilds the graph on the foreground subgraph and re-fits the likelihoods. That loop is never described, has no convergence criterion, and is not ablated. As written it is ambiguous: restricting to foreground nodes leaves no background seeds for p_B. If the reported gains come from that loop, the contribution is not the one-shot energy minimization the text claims. This is a reproducibility issue that needs to be fixed before the numbers can be trusted.\n\nOther issues are more minor but real: the multi-round protocol lets users choose \"informative views\" without a specified procedure, so the 92.2 mIoU is not independently reproducible; there are no error bars, so the 92.2 vs 92.5 gap is within noise; the Neyman-Pearson claim in Sec 3.4 overstates things because the densities are estimated, not known; and several hyperparameters (w_x, w_c, w_o, the coverage threshold, max iterations) are unreported. No code is released. The SH-ambiguity citation to the authors' own 2025 paper is fine if the claim holds, but it is a self-citation worth verifying.\n\nOverall: the core idea has merit and the paper is worth a serious referee, not a desk reject. The referee should demand a clear description and ablation of Algorithm 1, a fixed protocol for multi-round refinement, error bars, and code. If the authors address those, I would cite and build on it. As is, I would hold off citing because the actual algorithm is not fully specified.","headline":"A competent training-free graph-cut approach to sparse-scribble 3DGS selection, but the undocumented iterative loop in Algorithm 1 and loosely specified multi-round protocol make the headline numbers provisional until clarified.","tokens_in":12591,"tokens_out":6057,"would_cite":false,"duration_ms":53277,"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":"Training-free graph cuts on 3D Gaussian superpoints select an object from one scribble view at 85.3 mIoU.","keywords":["3D Gaussian Splatting","interactive object selection","graph cut","superpoints","scribble lifting","training-free segmentation","human-in-the-loop","visibility-aware lifting"],"falsifier":"On a single NVOS view, shift the foreground scribble by one pixel and rerun the single-round pipeline; if mIoU drops by more than a few points, the transmittance-weighted lifting in Eq. (6) is too fragile. In a custom scene with a transparent object in front of a similarly colored object, scribble only the front object from one view: the alpha-composited coverage will mix both surfaces, and if the resulting selection still achieves the reported ~85 mIoU regime, the lifting assumption survives; if not, the central claim is contradicted.","tokens_in":11579,"feed_emoji":"🎯","tokens_out":8201,"duration_ms":68759,"temperature":0.7,"pith_summary":"GaussianSelector aims to show that complete 3D object selection in a reconstructed Gaussian Splatting scene can be driven by very sparse user input—one or a few 2D scribble strokes—without retraining the scene representation and without any pretrained image segmentation model. It turns the native Gaussian primitives into a superpoint graph: appearance is reparameterized by a deterministic Canonical Axis Color descriptor, Gaussians are grouped by spatial and appearance proximity, and edges are weighted by continuity of position, color, and opacity. User scribbles are lifted into 3D by alpha-transmittance coverage, and the object is recovered as the global minimum of a binary graph-cut energy. On the NVOS benchmark the paper reports 85.3 mIoU from a single scribble view and 92.2 mIoU after two additional refinement rounds, close to the 92.5 of a strong all-views, SAM-mask method, while running in minutes on CPU without GPU inference. If these numbers hold, interactive 3D selection becomes practical where dense multi-view supervision and large GPU memory are unavailable.","feed_headline":"One scribble view selects 3D objects via graph cut at 85.3 mIoU","feed_subtitle":"Method matches multi-view SAM-based segmentation with sparse scribbles, no GPU, and no retraining.","key_machinery":"The load-bearing construction is the continuity-weighted superpoint graph combined with visibility-aware scribble lifting. The edge weight $w_{ij} = \\exp(-d_{ij}^2/\\sigma_{ij}^2)$ with $d_{ij} = w_x d_x + w_c d_c + w_o d_o$, self-tuned $\\sigma_{ij}$ from local k-NN distance medians, encodes where label boundaries should and should not fall; the scribble-lifting coverage $\\rho_i = \\sum_p \\alpha_{ip} T_{ip} M(p) / (\\sum_p \\alpha_{ip} T_{ip} + \\epsilon)$ maps 2D strokes onto Gaussian seeds through $\\alpha$-composited visibility. Together they reduce interactive 3D selection to a one-time cached graph construction plus a fast per-round min-cut, which is what makes the claimed CPU-only, multi-round","core_discovery":"GaussianSelector's central discovery is that a 3D object's identity is already encoded in the intrinsic structure of Gaussian primitives, so it can be recovered by a graph-cut labeled over scene-native superpoints, with no learned feature field or lifted SAM masks. Each Gaussian is described by an 18-dimensional Canonical Axis Color (CAC) feature—SH radiance evaluated along the six canonical axes of the primitive's anisotropic support—which avoids the non-unique representation ambiguity of raw SH coefficients. Gaussians are aggregated into superpoints via Leiden community detection on spatial and appearance proximity, and a k-NN superpoint graph carries continuity weights from normalized spa","pith_inferences":["If the Neyman-Pearson framing is taken literally, the method's ceiling is set by how well two GMMs capture true foreground/background appearance; replacing them with nonparametric density estimates would be a direct test of whether the 85.3 single-view number is appearance-model-limited.","Because CAC samples only the DC and first-order directional SH response, specular or view-dependent materials may be conflated; a testable extension is to evaluate on scenes with highlights or transparent objects, where Section 3.3's transmittance lifting is most fragile.","The same superpoint-graph-plus-visibility-lifting scaffold could be transferred to other primitive-based radiance representations or point clouds once an equivalent appearance descriptor and coverage map are defined; the paper does not claim this.","A likely practical consequence the paper leaves implicit: because the scene graph is cached once, the per-round update cost is small enough to support real-time interactive editing on a laptop CPU, which the reported 0.2-minute rounds suggest."],"forward_implications":["Single-view scribble interaction becomes a usable mode: the paper reports 85.3 mIoU from one NVOS scribble view, improving on the 70.1 of the original NVOS method.","A user can refine a selection round by round from new viewpoints, with each round reusing the cached graph and costing roughly 0.2 minutes on CPU; the paper reports 89.6 mIoU after one refinement round and 92.2 after two.","No pretrained segmentation network or GPU feature-field training is needed for the interaction itself, so the approach drops VRAM requirements and removes dependency on SAM-quality masks.","On 3D-OVS the method reaches 93.2 mIoU (93.6 with the ROI variant), within reach of the 94.4 reported by the all-views GaussianCut baseline.","The graph-cut formulation makes the result globally optimal for the given evidence and prior; the ablation shows scribbles alone yield 33.2 mIoU, graph alone 61.0, and the full model 85.3."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting primitive representation that the method operates on and whose renderer defines alpha-compositing for visibility lifting.","marker":"(Kerbl et al. 2023)"},{"why":"Supplies the NVOS benchmark, the single-view scribble interaction setting, and the 70.1 mIoU baseline that the single-round result must beat.","marker":"(Ren et al. 2022)"},{"why":"GaussianCut is the closest graph-cut-on-3DGS baseline and the strongest comparison (92.5 mIoU with all views and SAM masks).","marker":"(Jain, Mirzaei, and Gilitschenski 2024)"},{"why":"Provides the s-t min-cut algorithm that exactly minimizes the binary submodular Potts energy in Eq. (1).","marker":"(Boykov and Funka-Lea 2006)"},{"why":"Supplies the 3D-OVS benchmark used for the second quantitative evaluation.","marker":"(Liu et al. 2023)"},{"why":"Identifies the non-unique representation ambiguity of raw SH coefficients, motivating the CAC appearance descriptor.","marker":"(Xin et al. 2025)"}],"fun_headline_variants":["Scribble one view, graph-cut selects 3D objects, no training","Graph-cut selection from sparse scribbles at 85.3 mIoU","Sparse scribbles pick 3D objects via graph-cut, no GPU","Training-free 3D selection: graph-cut on Gaussian superpoints","One scribble view: graph-cut 3D object selection at 85.3 mIoU"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that a scribble's visible footprint in a rendered view, weighted by how much each Gaussian contributes to those pixels, reliably reveals the user's intended object; at object boundaries, with transparent or reflective surfaces, or when one Gaussian serves both scribbled and unscribbled pixels, wrong seed labels can be created that later stages cannot fully repair.","fun_headline_variants_meta":{"raw":{"variants":["Scribble one view, graph-cut selects 3D objects, no training","Graph-cut selection from sparse scribbles at 85.3 mIoU","Sparse scribbles pick 3D objects via graph-cut, no GPU","Training-free 3D selection: graph-cut on Gaussian superpoints","One scribble view: graph-cut 3D object selection at 85.3 mIoU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1112,"prompt_tokens":758,"completion_tokens":354,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":263}},"tokens_in":502,"tokens_out":354,"duration_ms":3930,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:06:04.063213+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a single NVOS view, shift the foreground scribble by one pixel and rerun the single-round pipeline; if mIoU drops by more than a few points, the transmittance-weighted lifting in Eq. (6) is too fragile. In a custom scene with a transparent object in front of a similarly colored object, scribble only the front object from one view: the alpha-composited coverage will mix both surfaces, and if the resulting selection still achieves the reported ~85 mIoU regime, the lifting assumption survives; if not, the central claim is contradicted.","supporting_citations":[],"review_version":1}