{"id":"a22f31b7-72d4-4002-9990-a0f781ee0d70","arxiv_id":"2607.27592","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A feedforward network trained only on 2D foundation-model features predicts rotation-robust, general-purpose 3D mesh features that work zero-shot for segmentation, correspondence, and deformation.","lead":"MeshFM trains a neural network to predict 3D surface features for meshes using only 2D image features from models like DINOv2, with no 3D annotations. The learned features transfer zero-shot to part segmentation, dense correspondence, and mesh deformation, and stay stable under arbitrary 3D rotations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training/evaluation overlap is not ruled out: PartObjaverse-Tiny, DenseCorr3D, and Manifold40 derive from Objaverse, and the paper never states that the 130k training shapes exclude these test sets.","rationale":"The reader identified the SAM aliasing-correction assumption as the weakest point. That is a real risk and deserves the proposed ablation. However, the most load-bearing condition for the central claim is that the feedforward network generalizes to shapes it has never seen. The paper trains on a large Objaverse subset, and several evaluation datasets are also Objaverse-derived. Without an explicit disjointness guarantee, the headline results could be explained by per-shape memorization in Stage 2 rather than by the quality of the distilled 2D features. This is not an accusation of misconduct; it is a missing control that is standard for zero-shot generalization claims. The paper's own two-stage design makes the risk concrete: Stage 1 produces per-shape teacher fields for training shapes, and Stage 2 regresses exactly those fields. If a benchmark shape was one of those training shapes, the model has already seen its teacher features during training. The SAM-correction issue affects the quality of the teacher, but even a perfect teacher cannot validate the generalization claim if the test shapes are not held out. I therefore keep the reader's CONDITIONAL verdict, with the added explicit condition that the authors demonstrate train/test disjointness. The concrete test above would settle this quickly and is a prerequisite for the stronger '2D features are all you need' framing.","tokens_in":16483,"tokens_out":6603,"duration_ms":57642,"concrete_test":"Request the complete Objaverse ID list used for Stage 1/Stage 2 training and the test IDs for PartObjaverse-Tiny, DenseCorr3D, Manifold40, and TOSCA. Compute the intersection. If non-empty, drop those IDs from training and re-run the downstream evaluations (or at least re-evaluate the released checkpoint on the disjoint subset); if the headline mIoU/AUC numbers shift by more than ~1 point/0.02, the generalization claim is not established. A quick prior check: use the released model to retrieve nearest training shapes for a few test shapes via mean-pooled features and inspect for near-duplicates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a feedforward network trained solely with 2D supervision generalizes to unseen 3D shapes. The paper trains on ~130k Objaverse shapes (S2.1) but never states that the evaluation benchmarks are disjoint from this training set. PartObjaverse-Tiny, DenseCorr3D, and Manifold40 are constructed from Objaverse/Objaverse-XL, so overlap is plausible. Because Stage 1 optimizes a per-shape teacher field and Stage 2 regresses those exact fields, any test shape included in training gives the network direct memorization access to its teacher features. In that case Tables 1-5 and S1-S2 would not demonstrate zero-shot transfer from 2D features to unseen geometry; they would only show that the regressor can reconstruct teacher fields for shapes it saw. This concern is more load-bearing than the SAM-correction assumption: even a perfect aliasing correction cannot establish generalization if the test shapes are in the training set. A clear statement of train/test ID disjointness, or a retrained evaluation on a held-out split, is a necessary condition for the paper's headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MeshFM, a two-stage framework for learning general-purpose 3D shape features from 2D foundation models. In Stage 1, per-shape neural feature fields are optimized by distilling DINOv2 (or other) features from multi-view renderings, with a SAM-based aliasing correction that replaces outlier pixels in each segment by the segment median. In Stage 2, a feedforward PVCNN+triplane-transformer network is trained to regress these teacher fields, with SO(3) rotation augmentation. The resulting features are then applied zero-shot to part segmentation, dense correspondence, classification, and deformation, and compared against PartField, Diff3F, and DenseMatcher. The central claim is that properly distilled 2D features are sufficient for near-SOTA 3D understanding without task-specific heads or 3D supervision.","tokens_in":16761,"tokens_out":4205,"duration_ms":39728,"significance":"If validated, the paper would support a strong and useful thesis: that 2D foundation features, when distilled with boundary-aware correction, can serve as a general-purpose representation for diverse 3D tasks, eliminating the need for task-specific 3D supervision. The paper is well positioned in the literature and includes extensive experiments, a fair common postprocessing protocol, ablations (SAM correction, foundation-model choice, two-stage design), inference-time analysis, and cost breakdowns. These are genuine strengths. However, the current evaluation has critical gaps: no demonstration of train/test disjointness for Objaverse-derived benchmarks, no error bars or multiple seeds, and a clear case where the paper's reported numbers contradict the 'on par' wording (TOSCA). These issues must be resolved before the central claim can be accepted.","major_comments":[{"comment":"The paper never states that the 130k-shape Objaverse training subset (S2.1) is disjoint from the evaluation benchmarks. PartObjaverse-Tiny (Table 1, S1) and DenseCorr3D (Table 3) are derived from Objaverse/Objaverse-XL, so train/test overlap is plausible. Because Stage 2 regresses the exact teacher fields produced in Stage 1, any test shape appearing in training could be memorized rather than generalized to. A clear statement of disjointness, or a retrained evaluation on a held-out split, is a necessary condition for the paper's headline claim of zero-shot transfer to unseen 3D shapes.","section":"S2.1, §4"},{"comment":"No error bars, confidence intervals, or multiple seeds are reported. The reported gains over PartField are very small (e.g., Table 1: 0.549 vs 0.542; Table 2: 0.520 vs 0.517), and in Table S1 MeshFM is actually below PartField on original PartObjaverse-Tiny instance segmentation (0.598 vs 0.652) despite the text saying 'slightly underperforms'. Without variance estimates, the claims of statistical superiority or parity are not supported. The authors should report mean and standard deviation over at least 3 runs, or otherwise justify that the differences are not noise.","section":"Tables 1, 2, S1, S2"},{"comment":"On the original (non-rotated) TOSCA benchmark, MeshFM reports AUC 0.34 and Err 18.8, while Diff3F reports 0.48/17.1 and DenseMatcher 0.50/10.2. The text states 'Our method is on par with state-of-the-art baselines,' but these numbers show a substantial underperformance, especially in AUC. This directly weakens the generality claim in the abstract and conclusion. The authors should either report additional metrics or revise the claim; if the underperformance is due to the functional-map postprocessing, that should be analyzed and disclosed.","section":"Table 4, §4.2"},{"comment":"The SAM aliasing correction assumes that each SAM segment has a unimodal feature distribution and that the median is a reliable representative. If a SAM mask is imprecise, or if patch bleeding corrupts the segment median itself, replacing outliers by the median can inject incorrect features into the teacher field, and every downstream result inherits this error. The ablation (Table 6, S4) shows only modest improvements (e.g., +0.031 mIoU average on PartObjaverse-Tiny; +0.01 AUC on DenseCorr3D), so it is not clear that this mechanism is the 'key enabler' described in Section 1. The authors should test robustness to SAM mask errors, e.g., by perturbing masks or by ablating with an alternative boundary correction.","section":"Eq. (1)-(2), §3.1"}],"minor_comments":[{"comment":"Typo: 'Championingthishypothesis' should be 'Championing this hypothesis'.","section":"§1"},{"comment":"The labels 'Original Distillation' and 'Corrected Distillation' are small; consider enlarging or adding arrows to clarify the comparison.","section":"Figure 3"},{"comment":"The threshold tau=1 is stated without sensitivity analysis. Since this is a hyperparameter of the proposed correction, a brief study across tau values would strengthen the paper.","section":"§3.1"},{"comment":"The sentence 'This preprocessing step is necessary regardless whether Stage 1 is incorporated' is unclear: for the Stage-2-only ablation, is the preprocessing identical? Clarify the comparison.","section":"S2.3"},{"comment":"Manifold40 classification: the paper reports accuracy but does not describe the number of training samples per class for the MLP, or the data split. Please provide details for reproducibility.","section":"§4.3"},{"comment":"The abstract states 'perform on par with methods trained explicitly with 3D supervision' without qualification; Table 4 shows a counterexample. Recommend adding a qualifying phrase such as 'on the evaluated benchmarks except TOSCA' or reworking the claim.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and important question, and the two-stage distillation idea is promising. However, the evaluation currently leaves open the possibility of train/test leakage for Objaverse-derived datasets, which is the most serious issue. If the authors can provide a clear disjointness statement or retrain on a held-out split, the paper may be acceptable after experimental strengthening. Also note that the small performance differences with no error bars will need attention from the editor; this is a typical issue in this area but should be fixed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MeshFM is a clean, well-engineered distillation pipeline, but the title-level claim is ahead of the evidence. The method itself is genuinely new and mostly sound; the gaps are in the evaluation and disclosure.\n\nWhat's new: a feedforward network that predicts per-shape optimized teacher fields, so you get DFD-quality features without test-time optimization. The SAM-based outlier replacement is a simple fix for patch bleeding and the ablation shows it helps. The SO(3) augmentation is straightforward and the robustness gains on rotated inputs are real. Inference speed is dramatic: 0.13s per shape versus minutes for Diff3F. The two-stage ablation supports the design.\n\nSoft spots, in rough order of importance. First, the paper never states that the Objaverse training subset is disjoint from the evaluation sets. PartObjaverse-Tiny and DenseCorr3D are built from Objaverse/Objaverse-XL; if any test shapes appear in the 130k training set, the feedforward regressor could memorize the teacher fields and the zero-shot claim collapses. This must be fixed with an explicit disjointness statement or a held-out split. (Manifold40, I believe, is from ModelNet, so that one is probably fine.)\n\nSecond, the segmentation wins over PartField are within noise: 0.549 vs 0.542 on PartObjaverse-Tiny, 0.520 vs 0.517 on PartNetE. No error bars or seeds are reported, so \"on par\" is the honest description, not \"outperforms.\" Third, the TOSCA original-pose result is clearly worse than Diff3F (AUC 0.34 vs 0.48), which the text glosses over. Fourth, the rotation comparison stacks the deck: baselines are evaluated without rotation augmentation; the paper even concedes PartField could be trained with it. That doesn't invalidate the result, but it weakens the \"handily outperforms\" framing.\n\nThe limitations section is honest about inherited 2D model biases, and the work is reproducible in principle (the authors should release code and checkpoints). The citation pattern looks fine.\n\nBottom line: this is a serious paper worth refereeing. An editor should send it out, with instructions that the authors clarify train/test disjointness, add variance estimates, and soften the \"all you need\" framing. For a reading group, it's a good discussion piece.","headline":"Solid distillation pipeline, but the 'all you need' claim needs a disjointness statement and more honest comparisons before it earns the title.","tokens_in":17286,"tokens_out":3142,"would_cite":true,"duration_ms":30115,"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":"A feedforward network trained purely on 2D supervision can match or beat task-specific 3D models on segmentation and correspondence, and stays accurate under arbitrary rotations.","keywords":["MeshFM","2D feature distillation","feature field","3D shape understanding","part segmentation","dense correspondence","rotation robustness","zero-shot transfer"],"falsifier":"Run the pipeline with the segmentation-based correction replaced by a comparable denoiser that does not depend on mask boundaries (for example, median filtering in feature space over local neighborhoods), or perturb the segmentation masks slightly on shapes with thin parts. If downstream segmentation and correspondence do not degrade when the masks are inaccurate, the specific mask-median mechanism is not the decisive factor; if the method collapses under corrupted masks, the assumption is confirmed as load-bearing.","tokens_in":16340,"feed_emoji":"🧊","tokens_out":4667,"duration_ms":41433,"temperature":0.7,"pith_summary":"The paper sets out to show that task-specific 3D networks are not necessary for 3D shape understanding: the apparent need for dedicated heads is, in the authors' view, an artifact of imperfect feature distillation. It introduces MeshFM, a feedforward network that predicts a general-purpose, rotation-robust 3D feature field for a mesh in one forward pass, trained only with 2D supervision. The enabling steps are a two-stage distillation—first fitting a clean per-shape teacher field from multi-view 2D features, with a segmentation-based correction that removes patch-induced 'feature bleeding' across part boundaries, then regressing that field with a point-cloud network under full rotation augmentation. If correct, the claim implies that a single set of 2D-derived features can replace specialized heads for segmentation, correspondence, classification, and deformation, and that 3D annotation is not a bottleneck for these tasks.","feed_headline":"2D-only features match 3D-supervised models on shape tasks","feed_subtitle":"MeshFM distills foundation-model features into rotation-robust 3D fields, beating task-specific baselines without 3D labels.","key_machinery":"The load-bearing object is the corrected teacher feature field: a per-shape neural field fitted by back-projecting multi-view 2D foundation-model features onto the mesh, then denoised by replacing outlier pixels within each segmentation mask with that mask's median feature. This clean field supervises a feedforward predictor (point-voxel encoder with triplane representation and transformer) that produces rotation-robust features in a single pass.","core_discovery":"The paper's central discovery is that distilled 2D foundation-model features, once corrected for aliasing and passed through a two-stage training procedure, are sufficient to match or beat task-specific 3D-supervised methods on part segmentation, dense correspondence, classification, and mesh deformation. The enabling correction is a segmentation-based feature-aliasing step: within each segment of an image-based segmentation model, pixels whose features deviate too far from the segment's median feature are replaced by that median, suppressing patch-level 'feature bleeding' across semantic boundaries. After the corrected teacher field is optimized, a feedforward network can regress it directl","pith_inferences":["If this claim holds, progress in 3D understanding could piggyback on progress in 2D foundation models: each new 2D model can be plugged into the same two-stage pipeline, a possibility the paper's foundation-model ablation already hints at.","The acknowledged left-right confusion in correspondence suggests a concrete extension: inject a weak geometric or positional signal into the corrected teacher field, which could remove the main failure mode while keeping the features annotation-free.","The fact that rotation robustness comes purely from augmentation suggests task-specific 3D models trained only on canonical poses could adopt the same training recipe, though their dependence on 3D-supervised or mask-based signals may make it harder.","A natural stress test of the thesis is scene-level understanding, where occlusion and multi-object fields introduce ambiguities that single-object distillation does not; the paper's per-object results leave this open."],"forward_implications":["On class-agnostic part segmentation, the general-purpose features reach or slightly exceed a leading segmentation-specialized method on the original test sets, and clearly outperform it when shapes are randomly rotated.","On dense correspondence, the method matches task-specific correspondence methods on canonical poses and remains accurate under SO(3) rotation, where the baselines degrade sharply.","The same features serve segmentation, correspondence, classification, and deformation zero-shot, whereas the task-specific baselines do not transfer across tasks.","The segmentation-based aliasing correction improves teacher-field quality consistently, raising instance-segmentation mean IoU by about three points on average and improving correspondence accuracy.","The two-stage design beats directly training the feedforward network on back-projected 2D features, indicating that teacher-field optimization filters multi-view inconsistency."],"fun_headline_variants":["2D features alone match 3D-supervised models on shape tasks","No 3D labels? Distilled 2D features still match supervised methods","MeshFM: 2D foundation features enable rotation-robust 3D understanding","Zero 3D supervision: 2D feature distillation rivals explicit 3D training","2D teacher features transfer to 3D, matching supervised baselines"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The correction assumes that pixels inside a single segmentation mask should all carry nearly the same feature and that large deviations are patch-bleeding noise; if the mask is imprecise or the bleeding contaminates the segment median itself, the corrected teacher field injects wrong features into every downstream prediction.","fun_headline_variants_meta":{"raw":{"variants":["2D features alone match 3D-supervised models on shape tasks","No 3D labels? Distilled 2D features still match supervised methods","MeshFM: 2D foundation features enable rotation-robust 3D understanding","Zero 3D supervision: 2D feature distillation rivals explicit 3D training","2D teacher features transfer to 3D, matching supervised baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000853,"raw_usage":{"total_tokens":3527,"prompt_tokens":707,"completion_tokens":2820,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":2716}},"tokens_in":451,"tokens_out":2820,"duration_ms":18410,"temperature":1.0,"reasoning_tokens":2716,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:02:29.296768+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline with the segmentation-based correction replaced by a comparable denoiser that does not depend on mask boundaries (for example, median filtering in feature space over local neighborhoods), or perturb the segmentation masks slightly on shapes with thin parts. If downstream segmentation and correspondence do not degrade when the masks are inaccurate, the specific mask-median mechanism is not the decisive factor; if the method collapses under corrupted masks, the assumption is confirmed as load-bearing.","supporting_citations":[],"review_version":1}