{"id":"83c805a3-e787-4090-8de5-9b19f0fe297a","arxiv_id":"2608.00767","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Hi-TOPS, a training-free meso-scale Flow-Freeze prior with TSDF-guided superquadric fitting, achieves competitive 3D part decomposition and the best mIoU on PartNet (55.87).","lead":"A new geometry-only algorithm, Hi-TOPS, cuts 3D meshes into meaningful parts by scoring curvature, normals, and density across several voxel scales, then using that score to guide superquadric shape fitting. It matches or beats many trained segmentation models on three benchmarks without any semantic labels or 2D foundation priors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Density factor √φ_den in Eq. 3 is face-count dependent; remeshing can move the top-15% Freeze set, so test tessellation stability before accepting the 'stable' prior claim.","rationale":"The paper makes two nested claims: (1) the S_B field is a reliable meso-scale proxy for part boundaries, and (2) the TSDF-constrained superquadric pipeline converts that proxy into competitive decompositions. Claim (2) is extensively ablated, and the pipeline is specified in unusual detail, which is a real strength. Claim (1) is the load-bearing foundation: if S_B is not stable under remeshing, then the 'hierarchical topology-aware prior' is not actually a shape-level property, and all downstream gains are contingent on the input mesh's tessellation. The reader's weakest assumption is in the same territory, but my concern sharpens it to a specific mechanism: φ_den = n_i with only a square-root compression. The reader emphasized 'dense surface detail'; I would phrase the risk as face-count dependence, which can occur even on smooth regions if they are subdivided. This is why agreement is partial. The evaluation-protocol concerns (self-selected subset, hyperparameters tuned on the benchmark) are real, but they affect the size of the reported margins, not the internal validity of the mechanism; the tessellation issue challenges the mechanism itself. The manuscript explicitly anticipates this sensitivity in Sec. 5, so the authors have already flagged the right risk; the missing piece is a quantitative test. If the proposed remeshing check passes, the central claim survives this attack and the remaining caveats are standard evaluation-hygiene issues. If it fails, the 'stable' and 'topology-aware' adjectives need qualification, and a tessellation-controlled benchmark would be necessary before strong claims can be credited. Since the reader already issued CONDITIONAL, my concern does not move that verdict category; it just supplies a concrete causal path behind the condition.","tokens_in":24402,"tokens_out":7588,"duration_ms":87510,"concrete_test":"Use the released code on 20 PartObjaverse-Tiny / PartNet shapes with non-uniform face density. For each shape, generate a geometry-preserving uniform remesh (target edge length ≈ median original edge length, e.g., via isotropic remeshing) and rerun only the scoring stage (Eqs. 3–5 at r=64): compute Dice(F_freeze_original, F_freeze_remesh). Then run the full pipeline and compare mIoU on the original and remeshed inputs. If average Dice < 0.95 or average mIoU shifts by more than 1.5 points, the score is tessellation-dependent and the 'stable topology-aware prior' claim does not hold as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the structural score S_B in Eq. 3 is a stable, shape-level boundary proxy. Appendix B defines φ_den = n_i, the number of per-face barycentric samples in a voxel. Since Nface=36 samples are placed per triangle, n_i is proportional to local triangle count, not surface area or geometric complexity. The square-root modulation compresses but does not remove this dependence: subdividing a patch 4× quadruples n_i in those voxels and doubles S_B for every voxel in the patch. Because Freeze is defined as the global top-15% of S_B (Eqs. 4–5), non-uniform remeshing re-ranks voxels and can move Freeze labels from true articulations onto smooth, densely tessellated regions. SQ seeding (Alg. A.2) and TSDF inflation treat B_freeze as a hard constraint, so this directly changes part boundaries. The paper's own limitation section (Sec. 5) admits 'local geometric cues may remain sensitive to tessellation and dense surface details,' but the effect is never quantified. The headline 'stable decompositions' and the low cross-dataset variance rest on the prior being a shape property rather than a mesh-tessellation property. This is a concrete, testable weakness, not a logical contradiction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hi-TOPS, an unsupervised geometry-only method for 3D mesh part decomposition. It voxelizes a normalized mesh at three resolutions, computes per-voxel scores from mean/std/gradient of curvature, normal variation, and a density factor, then thresholds the top 15% as 'Freeze' regions. A rule-based octree fusion combines the three resolutions into a hierarchical Flow-Freeze prior. Superquadrics are seeded via EDT maxima ('Body') and residual-component centroids ('Surface'), inflated by TSDF-guided least-squares fits that must stay inside Flow voxels, and transferred back to the mesh via curvature-aware edge primitives with distance-weighted voting. On PartObjaverse-Tiny, PartNet-selected, and HY3D-Bench, Hi-TOPS reports mIoU 51.63, 55.87, and 47.22, outperforming primitive-based baselines, competitive with learned models, and showing the lowest cross-dataset mIoU standard deviation (3.5).","tokens_in":24771,"tokens_out":7171,"duration_ms":85693,"significance":"If the results hold, Hi-TOPS is a significant contribution: it demonstrates that a hand-designed meso-scale geometric prior can rival learned part segmentation without training data or 2D foundation models, and it provides a detailed, reproducible pipeline (pseudocode and hyperparameters in appendices). The PartNet result (+10.3 mIoU over PartField) is particularly striking. The main caveats are that the core score's tessellation sensitivity and the test-subset hyperparameter search are not yet resolved, so the strength of the quantitative claims is conditional. The paper ships detailed algorithm pseudocode, a full hyperparameter table, and component ablations, which are strengths.","major_comments":[{"comment":"The 'stable prior' claim is undermined because the score's density factor sqrt(phi_den) = sqrt(n_i) is face-count dependent. Since Nface=36 samples are placed per face, n_i is proportional to local triangle count, not surface area or geometric complexity. A 4x non-uniform subdivision of a patch quadruples n_i and doubles S_B in those voxels; because Freeze is the global top-15% quantile (Eqs. 4-5), remeshing can re-rank voxels and move Freeze labels from articulations to densely tessellated smooth regions. Since SQ fitting uses B_freeze as a hard constraint (Alg. A.2, flow-zone acceptance) and the edge-primitive assignment preserves freeze boundaries, this directly changes the output decomposition. The paper's own limitation (Sec. 5) says the cues 'may remain sensitive to tessellation' but no experiment quantifies this. Please add a remeshing-stability experiment (e.g., identical shape a","section":"Eq. (3), App. B"},{"comment":"The reported configuration is not a single truly dataset-agnostic setting: alpha, lambda, R, Nface, and q were fixed, but Appendix D.2 says the cue weights were selected by grid search on a fixed 36-shape subset of PartObjaverse-Tiny, and Table 4 uses the same subset to select the resolution hierarchy. No validation split is documented, and the final PartObjaverse-Tiny numbers in Table 1 are computed on the full 200-shape set that includes this subset. This makes part of the reported advantage on PartObjaverse-Tiny reflect selection on the test benchmark. Please either (a) introduce an explicit train/validation/test split, reporting main results on a held-out test set, or (b) report performance on all benchmarks under several reasonable hyperparameter choices to show the conclusions are insensitive. Note also that Table 4 shows {32,64,128} (mIoU 57.66) beats the chosen {16,32,64} (56.56)","section":"App. D.1-D.2, Table 4"}],"minor_comments":[{"comment":"Equation numbering is inconsistent: Eq. 8 in the main text is the SQ-to-mesh assignment, but App. E.2 (B.3) refers to Eq. 8 as the parameter-update acceptance rule. Renumber to avoid ambiguity.","section":"Sec. 3.3 / App. E.2"},{"comment":"The row 'w/o TSDF Res=32' is placed under 'Geometric Cues', but removing TSDF is a fitting-stage ablation, not a geometric cue. Move it to the TSDF-related ablation group for clarity.","section":"Table 3"},{"comment":"The sentence 'The table has been reorganized for clarity (same setting in Table 3)' appears to be an editorial leftover. Rewrite to describe the sensitivity table directly.","section":"App. D.2"},{"comment":"The reported mIoU uses many-to-one greedy matching, which can be optimistic for over-segmented predictions. Please add a sensitivity check with Hungarian matching or report over/under-segmentation statistics.","section":"App. A / Table 1"},{"comment":"Minor typos: 'Hi-Tops' in App. A should be 'Hi-TOPS'; Fig. 2 caption has repetitive 'Normal Normal' formatting.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope for a geometry processing venue, and the technical contribution is well detailed. The two major comments are empirical and testable; I do not see a fundamental flaw that would require rejection, provided the authors can show the score is stable under remeshing and that the quantitative results are not an artifact of hyperparameter selection on the test subset. I would ask for a proper validation split or a perturbation analysis before relying on the headline cross-dataset consistency numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi-TOPS is the real deal as a training-free baseline: it beats primitive-only methods by a wide margin, lands competitive mIoU against learned models, and ships with enough detail to reimplement from the appendix. The new content is the multi-resolution Flow-Freeze fusion and the body-surface SQ seeding with TSDF inflation; neither is a conceptual earthquake, but combining known pieces into a working, unsupervised pipeline is legitimately useful.\n\nWhat it does well: the method is fully specified, the appendix gives the actual algorithm and hyperparameter table, the dataset-distribution analysis in Appendix C explains the cross-dataset behavior of the learned baselines, and the limitation section is honest. The single-thread CPU runtime of ~130s is a real cost but not a flaw.\n\nSoft spots, in proportion. First, the headline hyperparameters were chosen by grid search on a 36-shape subset of PartObjaverse-Tiny (App. D.2), and the same benchmark family is used as the main evaluation, with no separate validation split reported. That makes the exact margins over PartField and PartSAM less solid than the method itself. Adding a true validation split or full-benchmark numbers would fix the worst of this. Second, PartNet as used here is 100 hand-picked shapes, not the full benchmark; the +10 mIoU claim should be labeled accordingly. Third—and this one is testable—the density factor √φ_den in Eq. 3 is really √n_i, and n_i counts per-face barycentric samples per voxel, so it is proportional to local triangle count, not surface area. Subdivision without geometry change inflates S_B in that patch and, because Freeze is a global top-15% cut, can shift Freeze labels onto smooth densely-tessellated regions. The paper's own limitation paragraph says local cues may be tessellation-sensitive, but it never quantifies this. A remeshing-invariance experiment or switching to an area-based density would settle it. I don't think this is a load-bearing flaw—the method clearly produces good decompositions—but calling the prior 'stable' needs that experiment.\n\nWho's it for: anyone needing an editable, supervision-free decomposition for shape editing/abstraction, and anyone benchmarking training-free methods. The citation pattern is fine. Deserves a serious referee. My recommendation: send it to review with the validation-split and tessellation concerns as the main revision points.","headline":"Hi-TOPS is a genuinely useful training-free mesh decomposition pipeline, but its headline quantitative claims rest on hyperparameters tuned on an evaluation subset, and its boundary score has a face-count dependence that should be tested before calling the prior 'stable'.","tokens_in":25250,"tokens_out":2103,"would_cite":true,"duration_ms":25389,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A training-free, multi-resolution geometric score that tells 3D part growth where to freeze can segment meshes as accurately as learned models—and more consistently across datasets.","keywords":["3D part decomposition","part segmentation","Flow-Freeze prior","topology-aware scoring","superquadric fitting","TSDF-guided inflation","meso-scale voxelization","training-free shape analysis"],"falsifier":"Feed Hi-TOPS a mesh whose ground-truth part boundary lies in a region of flat, low-curvature surface with no normal discontinuity and uniform sampling density, e.g., a cylinder whose cap-to-body seam has been smoothed so no geometric cue fires. If no voxel at the seam enters the top-15% Freeze set, the decomposition merges cap and body, falsifying the claim that the score is a reliable boundary proxy.","tokens_in":24329,"feed_emoji":"🧩","tokens_out":11987,"duration_ms":125138,"temperature":0.7,"pith_summary":"Hi-TOPS sets out to show that 3D part decomposition does not require semantic labels or large vision models: the shape's own geometry, read at the right scale, already says where parts separate. The paper identifies a structural-scale mismatch—global methods bridge over joints, local methods fragment under noise—and argues the decisive evidence lives at a middle scale. It therefore voxelizes the mesh at three resolutions (16, 32, 64), computes per-cell scores from curvature, normal variation, and sampling density, and splits the volume into Flow regions where primitives may grow and Freeze regions that stop growth at articulations and thin parts. Superquadrics are then inflated under a TSDF alignment objective with the Flow-Freeze map as a hard constraint, and the result is mapped back to mesh faces through curvature-aware edge primitives. The reported outcome is that this untrained pipeline beats primitive-based baselines, is competitive with trained models, and has the lowest cross-dataset mean-IoU variation (3.5) of any method tested, with its best result on PartNet (55.87).","feed_headline":"Training-free 3D part splitter outdoes learned models on PartNet","feed_subtitle":"Using only the shape's own geometry, it keeps seams at joints and thin parts, with no training data or 2D priors.","key_machinery":"The Flow-Freeze prior is a scalar structural field on the active voxels of a mesh, S_B(i) = (λ1 μκ_i + λ2 σκ_i + λ3 Δκ_i + λ4 φ_norm,i) · sqrt(φ_den,i), whose top-α level set (α = 0.15) marks boundary-critical 'freeze' cells while the complement marks expandable 'flow' cells. The field is computed at resolutions {16, 32, 64} and fused by four octree rules, then used as a hard constraint in TSDF-guided superquadric inflation: it determines where body and surface seeds are placed, which expansions are accepted, and where growth stops so that articulation seams and thin attachments survive.","core_discovery":"The central claim is that a hand-built, fixed-parameter geometric prior can stand in for learned semantics in part decomposition. On a voxelized mesh the paper defines a per-cell score S_B(i) = (λ1 μκ_i + λ2 σκ_i + λ3 Δκ_i + λ4 φ_norm,i) · sqrt(φ_den,i), computes it at resolutions 16, 32, and 64, thresholds the top 15% as Freeze, and fuses the three maps coarse-to-fine so that fine Freeze evidence survives while coarse Flow support is retained. The Flow-Freeze partition is then used as a hard constraint on superquadric inflation: seeds are placed first at the farthest interior points (bodies) and then at residual surface components (thin parts), growth is driven by minimizing a TSDF discrepa","pith_inferences":["Beyond the paper: the Flow-Freeze field could be used as a free source of pseudo-labels to train or distill a fast network, since the prior itself requires no annotation.","Beyond the paper: because the paper reports that {32,64,128} beats {16,32,64} in accuracy at much higher cost, allocating fine voxels only where the score is locally uncertain would likely recover most of that gain at a fraction of the runtime.","Beyond the paper: the cue-to-boundary proxy could be stress-tested on deliberately nonuniform meshes and on shapes with smoothed, unarticulated seams; this would separate genuine geometric content from implementation choices.","Beyond the paper: replacing superquadrics with a more flexible primitive family, such as generalized cylinders or convex decompositions, while keeping the Flow-Freeze constraint should extend the method to highly non-convex parts."],"forward_implications":["Because it needs no training data or 2D priors, the pipeline can be applied to any new mesh category without retraining or dataset-specific tuning.","The output is a set of editable superquadrics plus residual surface patches, so the same decomposition doubles as a compact abstraction for editing, animation, and controlled generation.","The structural-scale analysis implies that methods working only at global or only at local scales will keep bridging joints or fragmenting surfaces; boundary evidence at the meso scale is the missing ingredient.","On articulation-heavy, long-tailed benchmarks, the geometry-only prior is more consistent than trained models, indicating that distribution shift rather than representation quality is what limits those models."],"supporting_citations":[{"why":"provides PartField, the learned feature-field baseline that Hi-TOPS outperforms on PartNet and uses to benchmark cross-dataset stability.","marker":"Liu et al. 2025"},{"why":"provides PartSAM, the strongest trained baseline on HY3D-Bench, defining the learned-model comparison.","marker":"Zhu et al. 2025"},{"why":"supplies the PartObjaverse-Tiny benchmark and the greedy-matching mIoU protocol used for all evaluations.","marker":"Yang et al. 2024"},{"why":"supplies PartNet, the articulation- and thin-structure-heavy benchmark where the method reports its best mIoU.","marker":"Mo et al. 2019"},{"why":"provides the HY3D-Bench generalization set used to test cross-dataset transfer.","marker":"Hunyuan3D et al. 2026"},{"why":"Marching-Primitives is the primitive-fitting baseline and the source of the Gaussian inlier weighting used inside the TSDF fitting kernel.","marker":"Liu et al. 2023a"},{"why":"the robust superquadric recovery work that establishes the SQ representation and fitting baseline for primitive-based comparison.","marker":"Liu et al. 2022"},{"why":"provides the voxelization library used to build the multi-resolution grids on which the score is computed.","marker":"Zhou et al. 2018"},{"why":"supplies the dihedral-angle sharp-edge idea that Hi-TOPS repurposes to over-segment the mesh into curvature-aware edge primitives for label transfer.","marker":"Chen et al. 2025"}],"fun_headline_variants":["Geometry alone beats trained 3D part decomposers","No training, no 2D priors: 3D part split from shape alone","Topology-aware geometry prior outperforms learned 3D part models","Zero-training 3D part splitter relies solely on geometry","Shape geometry alone predicts 3D part boundaries precisely"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a fixed score on coarse box-shaped cells—built from curvature, normal change, and sampling density, with the top 15 percent marked as freeze—reliably indicates where parts separate on arbitrary shapes; if a true boundary creates none of those cues, the method will miss it.","fun_headline_variants_meta":{"raw":{"variants":["Geometry alone beats trained 3D part decomposers","No training, no 2D priors: 3D part split from shape alone","Topology-aware geometry prior outperforms learned 3D part models","Zero-training 3D part splitter relies solely on geometry","Shape geometry alone predicts 3D part boundaries precisely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001236,"raw_usage":{"total_tokens":4904,"prompt_tokens":727,"completion_tokens":4177,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":4087}},"tokens_in":471,"tokens_out":4177,"duration_ms":29935,"temperature":1.0,"reasoning_tokens":4087,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:18:59.676119+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed Hi-TOPS a mesh whose ground-truth part boundary lies in a region of flat, low-curvature surface with no normal discontinuity and uniform sampling density, e.g., a cylinder whose cap-to-body seam has been smoothed so no geometric cue fires. If no voxel at the seam enters the top-15% Freeze set, the decomposition merges cap and body, falsifying the claim that the score is a reliable boundary proxy.","supporting_citations":[{"cited_title":", title =","cited_arxiv_id":null,"evidence_quote":"the robust superquadric recovery work that establishes the SQ representation and fitting baseline for primitive-based comparison."}],"review_version":1}