{"id":"80628c3d-968f-479e-8b00-8e6aff96a443","arxiv_id":"2505.24162","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Backprojected DINOv2 features from rendered views are matched in feature space to propose symmetry planes, which are then verified with Chamfer distance, yielding state-of-the-art training-free 3D symmetry detection.","lead":"A no-training method detects reflection symmetry planes in 3D models by borrowing visual features from DINOv2, a pretrained image model, and mapping them onto the mesh. On a standard ShapeNet benchmark it beats both classical geometry methods and specialized trained networks, improving the F-score by about 15 percent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core premise — symmetric counterparts are feature-space nearest neighbors — is validated only on pre-aligned objects; the promised random-rotation experiment is missing and could falsify the zero-shot claim.","rationale":"The paper is a coherent training-free pipeline with a clear geometric verification step, and the reported aligned-set results are internally consistent. The Chamfer verification and the 1-degree clustering provide a strong check against false positives, and the invariance analysis does show that symmetric pairs have lower feature distances than random pairs. However, the entire method hinges on the two-nearest-neighbor matching step: if symmetric counterparts are not the nearest neighbors in feature space, candidates are essentially random and the geometric filters cannot recover true symmetries. Section 4's E(O) metric is a necessary condition for invariance but not sufficient to establish that the top-2 neighbors are symmetric counterparts; a point's adjacent surface points can have very similar features due to local smoothness. The decisive missing evidence is the random-rotation experiment, which the paper explicitly says was generated but never reports. This is not an internal inconsistency, but it is a correctness risk: DINOv2's orientation bias could make the invariance property hold only for objects aligned with the training distribution of natural images. The reader identified essentially the same weakest assumption; the test-set contamination in parameter selection is a secondary concern and might be mitigated by the robustness across configurations shown in Table 2. The requested check is cheap and directly falsifiable, so the appropriate disposition remains conditional acceptance pending the missing experiment.","tokens_in":8633,"tokens_out":6225,"duration_ms":79791,"concrete_test":"Run the method with the configuration that gave the reported best result in Table 2 (FM10K-4R-FI, 42 or 86 views) on the randomly rotated version of the 1000-object ShapeNet test set described in Section 5.1, and compute F-score and SDE with the same plane-distance thresholds. If the F-score is within a few points of the aligned-set value (0.865), the invariance concern is resolved. If it drops materially, the zero-shot claim fails for unaligned objects. As a complementary diagnostic, on 50 rotated objects with ground-truth planes, compute for each sampled point the fraction of cases where the true symmetric counterpart appears among the two nearest feature neighbors; compare this hit rate against the same statistic on the aligned objects. If the rotated hit rate is near chance, the matching step is the point of failure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of state-of-the-art zero-shot symmetry detection rests on Eq. (8): DINOv2 features back-projected to a mesh are invariant under reflection, so that for each point the two nearest feature-space neighbors are its symmetric counterparts. Section 4 does not directly test this: Eq. (9) measures the distance from F(x) to F(x'), where x' is the nearest surface point to the reflected position T_P(x), not whether the top-2 nearest feature neighbors contain the symmetric counterpart. More importantly, all reported results (Tables 1 and 2) are on the pre-aligned ShapeNet test set, whose ground-truth symmetry planes coincide with coordinate planes. The paper itself states in Section 5.1 that a randomly rotated version of the test set was generated, yet no results on it are presented anywhere. DINOv2 is trained on natural images with a strong orientation bias; rendering with flat gray material and four in-plane rotations may not eliminate sensitivity to object orientation relative to the camera. If invariance degrades under rotation, nearest-neighbor matching would produce mostly arbitrary trios, candidate planes would fail the Chamfer threshold, and F-score would drop substantially. Because the method is advertised as training-free and zero-shot, robustness to object orientation is not a peripheral detail; it is the load-bearing assumption. The absence of the random-rotation result is therefore the single most important gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free, zero-shot method for detecting reflection symmetry planes in 3D meshes. The pipeline renders an object from multiple viewpoints (Fibonacci sphere sampling), extracts DINOv2 patch features, backprojects them onto the mesh, and uniformly samples 10,000 points with interpolated features. Symmetry detection then proceeds by finding, for each point, its two nearest neighbors in feature space; each triplet of points generates four candidate planes (three pair-midpoint planes and one trio plane). Candidate planes are filtered by distance to the object center, verified by a Chamfer distance between the original and reflected point sets, clustered by angular proximity, and ranked by confidence. The authors report state-of-the-art results on the ShapeNet test subset used in PRS-Net and E3Sym, with SDE 0.31e-4 and F-score 0.865, substantially improving on the previous best F-score of 0.753.","tokens_in":8916,"tokens_out":2993,"duration_ms":42617,"significance":"If the central invariance claim holds, this is a meaningful demonstration that frozen visual foundation models can solve a 3D geometric task without any training or fine-tuning. The method is simple, uses frozen DINOv2 features, and supplements feature-space matching with geometric Chamfer verification, which is a clean and reproducible design. The claimed performance gain of about 15% in F-score over prior specialized methods is striking and, if reproducible and robust to orientation, would be a valuable contribution to the symmetry-detection literature. The paper also makes a useful empirical study of how view count, rotation augmentation, and sampling density affect feature invariance. However, the current evidence is limited to pre-aligned objects, and the validation of the core invariance premise is indirect and partly circular, so the significance hinges on additional experiments.","major_comments":[{"comment":"The paper states in Section 5.1: 'we generate an alternative version of the test set by applying random rotations to the objects before evaluation.' Yet no results on this randomly rotated test set appear anywhere in the paper. This is load-bearing, not a peripheral detail: the core premise is that DINOv2 features, back-projected onto a 3D mesh, are invariant to reflection symmetry (Eq. 8), and the invariance analysis in Section 4 is performed on objects whose ground-truth symmetry planes coincide with coordinate planes. DINOv2 is trained on natural images with strong orientation priors, and the flat-gray rendering with four in-plane rotations may not remove sensitivity to object orientation relative to the camera. The advertised 'zero-shot' and 'training-free' claims would be substantially weakened if F-score drops on randomly rotated objects. A table reporting SDE and F-score on the randomly rotated test set, and ideally the invariance error of Eq. (9) under random rotations, is required to support the main claim.","section":"Section 5.1 and absence of reported results"},{"comment":"The invariance analysis in Eq. (9) measures the feature distance between a point x and the nearest surface point to its reflected position T_P(x), with the reflection plane P taken from ground truth. This is not the same as the detection assumption in Section 3.2, which is that the two nearest neighbors of a point in feature space are its symmetric counterparts. Eq. (9) does not test whether the true symmetric counterpart is among the top-2 nearest feature-space neighbors; a low value of E could coexist with a scenario where many other points have even closer feature distances. The paper should directly evaluate the detection premise, for example by reporting, for random points on objects with known symmetry, the fraction of points whose ground-truth symmetric counterpart is contained in the top-2 nearest-feature-neighbor set, both for aligned and randomly rotated objects.","section":"Section 4, Eqs. (8)-(9) and the detection hypothesis"},{"comment":"The invariance analysis in Section 4.2 selects the number of viewpoints, the rotation augmentation, and the sampling density using a fixed subset of 50 objects drawn from the same ShapeNet test distribution and using their ground-truth symmetry planes. The same test distribution is then used for the final evaluation in Table 1. This is a circularity risk: the configuration that minimizes Eq. (9) is chosen on objects whose symmetry planes are known and whose alignment matches the coordinate axes, and that configuration is then evaluated on the same distribution. To support the zero-shot claim, the authors should either use a disjoint held-out set for parameter analysis, or show that the final configuration is not sensitive to the choice of the 50-object subset (e.g., by reporting results with multiple random subsets and with the default configuration from the literature).","section":"Section 4.2 and circularity of parameter selection"},{"comment":"The paper claims in the introduction that the method 'requires no training data or parameter tuning,' but Section 4.4 explicitly identifies an 'ideal configuration' (10K points, Fibonacci-sampled viewpoints, 4 rotations) chosen from the invariance analysis, and the method also has fixed thresholds tau1=0.01, tau2=1 degree, and k=10. The absence of official training is clear, but the claim of 'no parameter tuning' is too strong given that the invariance analysis is exactly a parameter-selection experiment on the evaluation distribution. The authors should either soften this claim or demonstrate that the final performance is insensitive to these thresholds and to the configuration choices within a reasonable range.","section":"Introduction, Section 3.2, Section 4.4"}],"minor_comments":[{"comment":"In Eq. (9), the second occurrence of 'x' in the sum should be the reflected counterpart, likely denoted x' or x_hat, to avoid confusion with the outer summation variable x; the current notation is inconsistent with the definition of x as the nearest point to T_P(x).","section":"Equation (9)"},{"comment":"The caption states 'F-score (%)' but the reported values are fractions (e.g., 0.865), not percentages; please correct the caption or multiply the values by 100.","section":"Table 1 caption"},{"comment":"There is a typo in 'Diffussion' (should be 'Diffusion') and the sentence 'This threshold was chosen to be strict...' is missing a period in Section 3.2; minor copyediting is needed throughout.","section":"Section 2, Related Work"},{"comment":"The evaluation of the Langevin method is reported on only 350 of the 1000 objects with ad hoc clustering parameters (eps=0.02, MinPts=15); this should be stated more prominently in the main text or explicitly labeled as an incomplete comparison, so readers do not misinterpret the absence of Langevin from Table 1 as a head-to-head result.","section":"Appendix A"},{"comment":"The list of viewpoint counts in the multi-view rendering step includes 144 in the text ('[6,14,26,42,62,86,144]') but Figure 2 and Table 2 use 114 as the maximum; please reconcile this discrepancy.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong, simple idea and achieves impressive numbers on the standard pre-aligned benchmark. The main risk is that the result is an artifact of the test set being aligned with the coordinate axes, and the authors explicitly state they generated a randomly rotated test set but never report its results. That must be addressed before publication. I would also encourage the editor to ask for the code/data release mentioned in Section 6, since the method's reproducibility depends on the rendering and backprojection details. The related-work section is adequate, though the comparison with Langevin is incomplete as acknowledged in Appendix A."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2505.24162.\n\nThe genuinely new thing is the combination: backprojected DINOv2 features onto a mesh, then use feature-space nearest neighbors to propose symmetry planes, with geometric verification. Prior backprojection work went after keypoints or segmentation, and symmetry detectors either learned on data or used pure geometry. So this is a real application-level contribution, and it's simple enough that someone else could reproduce the pipeline from reading the paper (modulo code release).\n\nThe headline result—F-score 0.865 vs. 0.753 for E3Sym, SDE 0.31e-4—is strong and consistent across a sensible range of render counts and sampling densities. The invariance analysis on 50 objects is a useful sanity check, and the control condition (T4) does show that rotation augmentation helps via viewpoint diversity rather than just more images. The end-to-end numbers on the pre-aligned ShapeNet test set are believable; the method clearly does something right.\n\nNow the soft spots, in order of severity.\n\nFirst, the paper says it generated a randomly rotated version of the test set but never reports results on it. That's the big one. DINOv2 is trained on natural images with orientation bias; gray flat rendering and four in-plane rotations may not remove sensitivity to object pose. If invariance degrades under rotation, the whole nearest-neighbor premise falls apart, and the 'zero-shot' claim becomes 'zero-shot on pre-aligned objects.' The authors should either report those numbers or explicitly scope the claim. This is not a minor omission.\n\nSecond, the invariance experiment in Section 4 doesn't directly measure what the algorithm relies on. It computes feature distance from a point to the nearest surface point of its reflected coordinate, not whether the top-2 nearest feature neighbors actually contain the symmetric counterpart. Those are related but not identical. The end-to-end success gives indirect evidence, but a direct recall-at-2 measurement would be cheap to do and would strengthen the paper.\n\nThird, parameter selection uses ground-truth symmetry planes from a 50-object subset of the same 1000-object test set. It's hyperparameter tuning on the test set, not training; that's a mild circularity, but it should be acknowledged, and ideally the parameters should be fixed on a separate validation set or shown to be robust.\n\nCode and data aren't available yet, and there are no error bars. Those are standard for this subfield, but they do limit how far I'd trust the single-run numbers.\n\nBottom line: this is a solid application paper, the idea is worth taking seriously, and the claimed improvement is likely real on the standard benchmark. The missing random-rotation experiment is the load-bearing gap. I'd send it to a serious referee, with a note asking the authors to supply that experiment before acceptance.","headline":"Training-free DINOv2 backprojection for 3D symmetry is a genuinely new trick with strong numbers, but the missing random-rotation experiment leaves the zero-shot claim under-supported.","tokens_in":9432,"tokens_out":2489,"would_cite":false,"duration_ms":28456,"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 training-free, zero-shot pipeline detects reflection-symmetry planes in 3D meshes by matching back-projected visual features, reporting F-score 0.865 versus 0.753 for the best prior method.","keywords":["3D symmetry detection","zero-shot","feature backprojection","DINOv2","foundation vision models","reflection symmetry","Chamfer distance","multi-view rendering"],"falsifier":"Randomly rotate the evaluation objects before detection and rerun the method: if the F-score drops toward the geometric baselines or the fraction of nearest-neighbor trios whose candidate planes align with ground truth collapses, the claimed feature invariance depends on canonical axis alignment rather than on the object's symmetry itself.","tokens_in":8421,"feed_emoji":"🪞","tokens_out":5609,"duration_ms":58072,"temperature":0.7,"pith_summary":"Symmetry detection in 3D meshes usually relies on geometric descriptors or learned networks. This paper argues that visual features taken from a pretrained vision transformer and mapped back onto the mesh surface already encode symmetry: a point and its mirror point receive nearly the same feature vector. The method therefore needs no training: it finds, for each surface point, its two nearest neighbors in feature space, forms candidate reflection planes from those trios, and keeps planes that pass near the object's center and whose reflected Chamfer distance is small. On a thousand-object test set the paper reports an F-score of 0.865, about fifteen percent better than the strongest prior approach, with the lowest reported symmetry distance error.","feed_headline":"Training-free method finds 3D symmetry planes, beating trained rivals","feed_subtitle":"Back-projected visual features match mirror points, lifting F-score to 0.865 with zero training data.","key_machinery":"Feature backprojection combined with two-nearest-neighbor trios. The object is rendered from multiple viewpoints; patch features from the pretrained vision transformer are averaged across visible views and barycentrically interpolated onto 10,000 sampled surface points, producing a 384-dimensional descriptor per point. In feature space, the nearest neighbors act as presumed symmetric partners, and trios of such partners generate candidate reflection planes that are then filtered, verified by Chamfer distance after reflection, and clustered by angular proximity.","core_discovery":"The central claim is that back-projected visual features from a pretrained vision transformer are approximately invariant under an object's reflection symmetry, and that this invariance can be converted directly into a symmetry-detection algorithm. For each point on a uniformly sampled set of 10,000 surface points, features are computed by rendering the object from many Fibonacci-sampled viewpoints, rotating each view four ways, averaging patch features over visible views, and interpolating within mesh faces. The algorithm then assumes that the two nearest neighbors of every point in L1 feature space are its symmetric counterparts. Each point trio yields four candidate planes—three bisector planes of point pairs and one plane through the trio—which are filtered to planes within five percent of the object's diagonal from the origin, verified by Chamfer distance between the point set and its reflection, clustered within one degree, and ranked by confidence. The paper reports F-score 0.865 and SDE 0.31e-4 on the evaluation subset, exceeding both geometric and learned baselines without any training.","pith_inferences":["If the feature invariance holds under arbitrary object pose, the same two-nearest-neighbor scheme could extend to partial or approximate symmetries by relaxing the origin-distance and Chamfer thresholds.","The paper tests objects that are pre-aligned with coordinate axes but reports no random-rotation results; arbitrary pose is the natural stress test for whether the invariance is intrinsic or relies on canonical alignment.","The reflection-matching idea could generalize to rotational or helical symmetries by grouping more than two nearest neighbors into orbits and fitting the corresponding transformation instead of a plane."],"forward_implications":["Zero-shot 3D symmetry detection becomes feasible without training data: the same pretrained features generalize directly to unseen meshes.","The confidence score and Chamfer verification step give downstream reconstruction, compression, or completion pipelines a ranked set of reliable global symmetry planes.","The method needs only a mesh and a frozen vision transformer, so it can be applied to new object classes without dataset-specific tuning.","Even the worst configuration tested, six viewpoints with no surface resampling, stays competitive, suggesting that the plane-verification stage carries much of the accuracy.","The approach is a direct demonstration that 2D foundation model features, once back-projected, can solve a purely geometric 3D task without any geometric supervision."],"supporting_citations":[{"why":"Supplies the pretrained vision transformer whose patch features are backprojected onto the mesh.","marker":"[12]"},{"why":"Provides the register-token variant of that transformer used for robustness to noise and high-frequency content.","marker":"[17]"},{"why":"Defines the evaluation dataset, the SDE and F-score metrics, and the strongest learned baseline that this method reports beating.","marker":"[9]"},{"why":"Supplies the four-rotation image augmentation strategy and an additional self-supervised baseline for comparison.","marker":"[10]"},{"why":"Provides the Fibonacci spiral viewpoint sampling that gives near-uniform coverage of the viewing sphere.","marker":"[16]"},{"why":"Demonstrates the feature backprojection technique for 3D keypoint detection that this work adapts to symmetry detection.","marker":"[14]"},{"why":"Shows backprojection of 2D visual features used for 3D part segmentation, supporting the transferability of backprojected features.","marker":"[15]"}],"fun_headline_variants":["Backprojected features find 3D symmetry planes, zero training","Training-free 3D symmetry via backprojected visual features","Zero-shot 3D symmetry from vision features, no training","Backprojected vision features reveal 3D mirrors without training","Foundation model features pinpoint 3D symmetry with zero training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the back-projected visual features of a point and its mirror-image point are closer to each other in feature space than to any other point on the surface, so nearest-neighbor matching really identifies symmetric counterparts.","fun_headline_variants_meta":{"raw":{"variants":["Backprojected features find 3D symmetry planes, zero training","Training-free 3D symmetry via backprojected visual features","Zero-shot 3D symmetry from vision features, no training","Backprojected vision features reveal 3D mirrors without training","Foundation model features pinpoint 3D symmetry with zero training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001873,"raw_usage":{"total_tokens":7302,"prompt_tokens":852,"completion_tokens":6450,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":6364}},"tokens_in":468,"tokens_out":6450,"duration_ms":40872,"temperature":1.0,"reasoning_tokens":6364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:31:54.105327+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Randomly rotate the evaluation objects before detection and rerun the method: if the F-score drops toward the geometric baselines or the fraction of nearest-neighbor trios whose candidate planes align with ground truth collapses, the claimed feature invariance depends on canonical axis alignment rather than on the object's symmetry itself.","supporting_citations":[{"cited_title":"Vision transformers need registers, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the register-token variant of that transformer used for robustness to noise and high-frequency content."},{"cited_title":"E3sym: Leveraging e(3) invariance for unsupervised 3d planar reflective symmetry detection","cited_arxiv_id":null,"evidence_quote":"Defines the evaluation dataset, the SDE and F-score metrics, and the strongest learned baseline that this method reports beating."},{"cited_title":"A dataset-free approach for self-supervised learning of 3d reflectional symmetries, 2025","cited_arxiv_id":null,"evidence_quote":"Supplies the four-rotation image augmentation strategy and an additional self-supervised baseline for comparison."},{"cited_title":"Measurement of areas on a sphere using fibonacci and latitude–longitude lattices.Mathematical Geosciences, 42:49–64, 2009","cited_arxiv_id":null,"evidence_quote":"Provides the Fibonacci spiral viewpoint sampling that gives near-uniform coverage of the viewing sphere."},{"cited_title":"Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features, 2024","cited_arxiv_id":null,"evidence_quote":"Demonstrates the feature backprojection technique for 3D keypoint detection that this work adapts to symmetry detection."},{"cited_title":"3d part segmentation via geometric aggregation of 2d visual features.IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025","cited_arxiv_id":null,"evidence_quote":"Shows backprojection of 2D visual features used for 3D part segmentation, supporting the transferability of backprojected features."}],"review_version":1}