{"id":"0c71bce5-b5c2-450d-b858-3afeed057f71","arxiv_id":"2411.17763","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A zero-shot symmetry detector built on DINOv2 and multi-view diffusion sets a new state of the art for single-image 3D symmetry detection and improves single-image 3D generation when integrated into DreamGaussian.","lead":"The paper introduces Reflect3D, a system that detects 3D reflection symmetry from a single photo and then uses that symmetry to improve single-image 3D generation. It reports large gains in symmetry detection accuracy and better geometry and texture for generated objects, making symmetry a practical and powerful structural prior for 3D content creation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth protocol assumes every symmetry plane passes through the bounding-sphere center; if that center is not the minimal enclosing sphere, true off-center symmetry planes are excluded, biasing training and evaluation.","rationale":"The reader's weakest assumption is the GT labeling protocol, and I agree it is load-bearing. My concern sharpens one unstated geometric condition: the protocol relies on the bounding-sphere center lying on every symmetry plane. Without knowing whether the sphere is minimal, the claim in Sec. 9.1 is not guaranteed. The same protocol is used for training and evaluation, so any systematic exclusion of off-center symmetry planes would put Reflect3D at an advantage over NeRD, which was trained on different labels. A concrete code-level check and re-evaluation with an independent centering would settle the matter. The paper's other contributions—the feed-forward detector design, the multi-view aggregation, and the symmetry-conditioned generation pipeline—are not themselves undermined by this concern; the issue is specifically the evaluation protocol. For this reason I maintain the reader's CONDITIONAL verdict: acceptance should require releasing the preprocessing code and confirming the centering assumption, or re-running the evaluation with a corrected protocol. No ad hominem; the concern is about an underspecified preprocessing step, not about intent.","tokens_in":12589,"tokens_out":17655,"duration_ms":162836,"concrete_test":"Release the preprocessing code and check whether the 'bounding sphere' is computed via a minimal-enclosing-sphere algorithm (e.g., Welzl's) or from the axis-aligned bounding box. If it is bbox-derived, regenerate ground truth for the GSO and OmniObject3D test sets using the minimal enclosing sphere center (or the object centroid), rerun the Table 2 evaluation for Reflect3D, Reflect3D-FF, NeRD, and Shape to Symmetry, and compare F@5 and GD. If the ranking and margins persist, the concern is resolved; if the GT plane count or the relative ordering changes materially, the reported state-of-the-art claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Appendix (Sec. 9.1) states that centering each mesh by its 'bounding sphere center' ensures the center of any ground-truth symmetry plane is at the origin. This is only mathematically guaranteed if the sphere is the unique minimal enclosing sphere; for a reflection-symmetric shape, the minimal enclosing sphere is invariant under the reflection, so its center lies on the mirror plane. If the implementation instead derives the sphere from the axis-aligned bounding box (the common convention for 'bounding sphere'), the center need not lie on the mirror plane. For example, a set of points symmetric about the line y=2x can have an axis-aligned bounding-box center off that line. Candidate planes are then constrained to pass through a point not on the true symmetry plane, so the Chamfer distance between the original and reflected point clouds is large, the candidate fails the manually selected threshold, and the true symmetry plane is never refined by ICP. The same protocol generates training labels (Objaverse+ShapeNet) and test labels (GSO, OmniObject3D); both the detector and the evaluation metrics are thus aligned to a potentially incomplete and biased notion of symmetry. Because Reflect3D is trained on this exact labeler while NeRD was not, the reported F-scores could overstate Reflect3D's advantage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Reflect3D proposes a transformer-based feed-forward detector for 3D reflection symmetry from a single RGB image, trained on a newly curated large-scale dataset (Objaverse+ShapeNet) with automatic Chamfer/ICP-derived symmetry labels. The detector uses a frozen DINOv2 encoder and cross-attention over 31 symmetry hypotheses. To resolve single-view ambiguity, the paper generates multiple views with a multi-view diffusion model, filters them by CLIP similarity, and aggregates per-view normal predictions via K-means clustering. The method is evaluated on GSO and OmniObject3D, reporting large improvements over NeRD and other baselines, and is integrated into DreamGaussian through symmetry alignment, symmetric SDS, symmetric densification, and symmetric texture refinement, yielding better CLIP-Sim and Chamfer distance. The paper positions the work as a scalable, zero-shot detector and demonstrates downstream utility in single-image 3D generation.","tokens_in":12884,"tokens_out":5416,"duration_ms":52076,"significance":"If the reported results hold, Reflect3D is a substantial advance in single-image 3D symmetry detection, showing that a scaled transformer trained on diverse synthetic data can generalize to real-world scans, and that the detected symmetry genuinely improves optimization-based 3D generation. The curated dataset with many more categories and symmetry planes than prior work is a valuable contribution. The paper also makes a good methodological point about using generative multi-view priors to reduce single-view ambiguity. The main risk lies in the automatic labeling protocol, which is the foundation for both training and evaluation; if that protocol has a systematic bias, the reported margins over baselines could be partly an artifact of shared label generation.","major_comments":[{"comment":"The ground-truth generation protocol centers each mesh by its \"bounding sphere center\" and then restricts all candidate symmetry planes to pass through the origin, claiming this \"ensures the center of any potential ground truth symmetry plane is on the origin.\" This guarantee holds only if the bounding sphere is the minimal enclosing sphere. For a reflection-symmetric shape, the minimal enclosing sphere is unique and its center lies on the mirror plane, but for an arbitrary bounding sphere (e.g., one derived from the axis-aligned bounding box or a Ritter-style approximation) the center need not lie on the mirror plane. In that case, candidate planes constrained to pass through the origin will have large Chamfer distances to the true plane, fail the manually selected threshold, and never be refined by ICP. Since the same protocol generates both the training labels (Objaverse+ShapeNet) and the test labels (GSO, OmniObject3D), this potential systematic omission directly affects all F-scores and geodesic distances in Tables 2 and 4. Please specify the exact bounding-sphere algorithm and, if it is not the minimal enclosing sphere, replace it with the minimal enclosing sphere or provide evidence that the resulting label set does not miss off-center symmetry planes.","section":"Appendix 9.1"},{"comment":"The problem definition in Eq. (3) states that a symmetry plane is parameterized by both its normal n_p and its distance d_p, but the proposed detector predicts only the normal and the evaluation metrics (geodesic distance, F@threshold) are computed solely on normals. The reported results therefore do not assess the full plane detection problem, and the abstract's claim of \"3D reflection symmetry detection\" is stronger than what is measured. Please either evaluate the full plane parameters (for instance, by aligning the predicted plane to the mesh and measuring offset error), or clearly and prominently state that the method and evaluation are for normal-only detection, and discuss the implications for downstream applications that require the plane position.","section":"Section 3.1 and Section 5"},{"comment":"No variance estimates, confidence intervals, or significance tests are reported for any quantitative result. The evaluation sets are finite (572 GSO objects and 100 OmniObject3D objects), and the improvements over the best baseline are large, so the main conclusions are likely robust, but the paper should include bootstrap confidence intervals or repeated-run standard deviations to support the state-of-the-art claim and to let readers judge the stability of the F-scores, especially at the tight F@5 threshold where the absolute numbers are small.","section":"Tables 2 and 3"},{"comment":"The manual selection of the Chamfer-distance threshold and the manual inspection of a \"random subset\" do not rule out systematic false negatives in the labeling protocol. In particular, if the threshold is set too strictly, the protocol will label as asymmetric objects that actually have a symmetry plane whose Chamfer distance is above the threshold due to surface noise or imperfect mesh quality. Since the detector is trained against these labels, a strict threshold would penalize the detector for predicting true planes, and the reported recall values could be depressed for all methods, while the training/evaluation alignment would still favor a model trained on the same threshold. Please report the threshold value, the distribution of Chamfer distances for accepted and rejected candidates, and the results of a more systematic human evaluation on a stratified sample of both positive and negative labels.","section":"Section 4 and Appendix 9.1"}],"minor_comments":[{"comment":"The reflection matrix has a formatting error: the top-right entry appears as \"−2dpnp\" without the proper row/column separation, and the matrix is not cleanly typeset. Please fix the LaTeX so that the 3x1 block and the 1x3 zero row are clear.","section":"Equation (2)"},{"comment":"In the row for the proposed dataset, the dataset name is written as \"Objaverse+ ShapeNet\" with an odd spacing, and the footnote marker for \"Ours\" is not defined. Please clean up the table formatting and explain the asterisk in the caption or text.","section":"Table 1"},{"comment":"The K-means clustering threshold is set to 30 degrees in the appendix, but no sensitivity analysis is provided. A short plot or paragraph showing the effect of varying this threshold on F-scores would help justify the choice and the claim that the method is not sensitive to it.","section":"Section 3.3"},{"comment":"The ablation study for the single-image 3D generation pipeline is qualitative only. Since Table 3 reports quantitative metrics, it would strengthen the paper to include a small table with CHD, CLIP-Sim, and F-scores for each ablated variant, so readers can see the magnitude of each component's contribution.","section":"Figure 4"},{"comment":"The term \"zero-shot\" is used to describe the detector, but the detector is trained on a large dataset of 84,789 objects. Please clarify in the abstract or introduction that \"zero-shot\" refers to generalization to unseen object categories and real-world scans without fine-tuning, not to a training-free method.","section":"Abstract and Section 1"},{"comment":"The paper states that the dataset has \"more comprehensive coverage of symmetry planes\" based on the higher number of symmetry planes per object, but it does not report the distribution of the number of symmetry planes per object. Adding a histogram or at least mean/median values would make this claim more transparent.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is timely and the empirical gains are large, but the correctness of the automatic labeling protocol is the linchpin of the entire evaluation, and the current manuscript does not provide enough detail to rule out a systematic bias in the ground-truth planes. The bounding-sphere-centering issue raised in the stress-test note is a concrete and plausible failure mode that the authors must address. I would ask the editor to insist on a precise description of the bounding-sphere computation, an analysis of the threshold and its consequences, and either an evaluation that includes plane offsets or an explicit reframing of the task as normal-only detection. If these points are resolved, the paper could be a strong contribution. If the authors cannot show that their bounding sphere is minimal, the reported F-scores may need to be recomputed on a corrected label set."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Reflect3D is the first single-image symmetry detector I'd actually trust on arbitrary objects. The frozen-DINOv2-plus-transformer-decoder over discretized hypotheses is simple and it works; the reported F@5 jump from 0.191 to 0.390 on GSO with multi-view aggregation is large and consistent with the ablations. The new Objaverse+ShapeNet dataset is a real resource. The integration into DreamGaussian shows a genuine downstream benefit, though the gains there are more modest.\n\nThe soft spots are in the labeler, not the architecture. Ground truth comes from a Chamfer-distance threshold plus ICP refinement, with the threshold manually selected. That's defensible, but the appendix's centering claim needs scrutiny: it says centering by the 'bounding sphere center' guarantees symmetry planes pass through the origin. That is only true for the minimal enclosing sphere. If the implementation uses an axis-aligned bounding-box sphere, as is common, then off-center symmetry planes are systematically excluded from both training and evaluation, and the detector and the metrics are aligned to the same bias. NeRD, trained on different labels, would then be measured against a labeler it never saw, which would inflate the apparent gap. This is not a fatal flaw, but the authors must state exactly what 'bounding sphere' means and, ideally, validate a subset against an independent protocol or human labels.\n\nThe other weakness is statistical: no error bars anywhere, and the generation comparison in Table 3 is a single run. The improvements are plausible but could be noise. A few seeds and a sensitivity analysis on the Chamfer threshold would firm it up.\n\nOn credit: the work is honestly presented, the ablations isolate components, and the limitations section is straightforward. The dataset and code release should be a condition, not a courtesy. Overall the central argument holds up: scale plus a good backbone plus multi-view diffusion resolves much of the single-view ambiguity.\n\nThis paper is for anyone working in symmetry detection, single-image 3D, or symmetry priors in generation. I'd bring it to a reading group. Recommendation: send it to a serious venue; conditional acceptance with requests for labeler clarification, threshold sensitivity, and error bars. It deserves careful refereeing.","headline":"Reflect3D is a legit step forward in single-image symmetry detection, but the auto-labeling protocol needs a sharper look before the absolute numbers should be trusted.","tokens_in":13375,"tokens_out":2948,"would_cite":true,"duration_ms":27623,"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":"Reflect3D detects an object's 3D mirror planes from a single photo, and the detected planes measurably sharpen single-image 3D generation.","keywords":["3D reflection symmetry","single-image symmetry detection","zero-shot learning","multi-view diffusion","single-image 3D generation","score distillation sampling","3D Gaussian splatting","Objaverse"],"falsifier":"Take symmetric CAD models with known, explicit mirror planes, render single images, and run both the automatic labeling protocol and Reflect3D on them. If the labeling protocol's planes disagree with the CAD-defined planes by more than a few degrees, or if the detector's error against the CAD-defined planes is much larger than the reported average geodesic distance (about $13^\\circ$ on GSO), then the benchmark labels — not the detector — are carrying the reported accuracy.","tokens_in":12428,"feed_emoji":"🪞","tokens_out":9307,"duration_ms":73053,"temperature":0.7,"pith_summary":"The paper claims that 3D reflection symmetry — the set of mirror planes through an object — can be recovered from a single RGB image by a transformer trained at scale, without category-specific priors or explicit 3D input at test time. It introduces Reflect3D, a zero-shot detector whose feed-forward pass alone already beats prior single-image symmetry methods on real scanned-object benchmarks, and whose multi-view variant (aggregating predictions from several diffusion-generated views) roughly doubles the top accuracy. The same predicted planes are then injected into a diffusion-based single-image 3D generator, improving both appearance (CLIP similarity) and geometry (Chamfer distance and F-scores) on the same test objects. If correct, symmetry becomes a cheap, reusable structural prior for making single-image 3D reconstruction and generation more faithful, especially for the unseen back side of objects.","feed_headline":"Mirror planes found from one photo sharpen 3D generation","feed_subtitle":"New symmetry detector outperforms prior zero-shot baselines and improves single-image 3D generation.","key_machinery":"The mechanism that carries the argument is a discrete set of symmetry hypotheses serving as query tokens. The detector enforces a hemisphere of $N=31$ fixed unit normals; each hypothesis is a small neighborhood of possible plane normals. A shallow MLP embeds these hypotheses into query features, cross-attention reads the DINOv2 image features per query, and one MLP head classifies whether each neighborhood contains a ground-truth plane while another regresses an exact normal as a quaternion residual. Multi-view aggregation is the second half: a multi-view diffusion model yields eight surrounding views, CLIP similarity filters inconsistent ones, predictions are rotated back to the input frame, and K-Means clustering with a $30^\\circ$ threshold outputs the final normals. For generation, symmetry enters DreamGaussian in three interventions—aligning the plane to a coarse Gaussian cloud via ICP, sampling SDS losses on mirrored views, and periodically reflecting 50% of Gaussians to densify symmetrically—plus a symmetric texture refinement stage that uses the mirrored image for back-side UVs.","core_discovery":"On the paper's own terms, the central discovery is that symmetry detection from a single image is no longer a category-restricted special case: a feed-forward transformer with a frozen self-supervised encoder and a cross-attention decoder over discrete normal hypotheses can predict mirror planes zero-shot, and a generative multi-view step removes the single-view ambiguity that previously forced methods onto narrower domains. Quantitatively, Reflect3D-FF lifts $F@5^\\circ$ from $0.040$ (NeRD) to $0.191$ on GSO, and with multi-view aggregation the same metric reaches $0.390$ while average geodesic error drops from $36.2^\\circ$ to $13.3^\\circ$. When the detected symmetry is used to guide DreamGaussian's SDS optimization — via plane alignment, symmetric SDS, densification, and symmetric texture refinement — CLIP-Sim rises by $0.037$ and Chamfer distance falls by $0.028$ on GSO, evidence that the symmetry prior transfers from detection into generation.","pith_inferences":["Because the detector only outputs the plane normal (not the plane distance), integrating it into a different 3D representation — e.g., a NeRF or a mesh from a feed-forward model — would require the same alignment trick used here, so the method's benefit in generation is likely tied to optimization-based pipelines that can afford an ICP alignment step.","The 31-hypothesis discretization could be made adaptive: a learned set of queries, or a coarse-to-fine cascade over normals, might remove the need for the multi-view diffusion stage entirely on objects with strong symmetry evidence.","If the automatic ground-truth protocol is perceptually aligned, the resulting dataset of ~85k objects with labeled symmetry planes could serve as pretraining data for other symmetry-aware tasks, such as symmetry-constrained pose estimation or shape completion; the paper only demonstrates detection and generation.","The success of a frozen DINOv2 suggests that symmetry planes are implicitly encoded in self-supervised visual features; probing which attention heads encode mirror relations could lead to a cheaper detector or to inserting symmetry biases into other vision transformers."],"forward_implications":["Symmetry detection transfers zero-shot from synthetic training to scanned real objects (GSO, OmniObject3D), so the method does not need category-specific retraining for new objects.","Single-view ambiguity is mitigated by generative multi-view synthesis and clustering, suggesting that ambiguous 3D cues can be resolved by leveraging diffusion priors rather than stronger 3D supervision.","A predicted symmetry plane is accurate enough to improve SDS-based 3D generation in both 2D appearance and 3D geometry, measured by CLIP-Sim, Chamfer distance, and F-scores.","The symmetric densification and mirrored SDS allow the pipeline to tolerate asymmetric details, so approximately symmetric real objects still benefit from the prior.","Freezing the DINOv2 encoder is critical: fine-tuning it degrades detection drastically, implying the pretrained self-supervised features already carry the geometric information needed for symmetry."],"supporting_citations":[{"why":"Frozen DINOv2 encoder that supplies the geometry-aware image features the symmetry decoder attends to.","marker":"[26]"},{"why":"Multi-view diffusion model used to synthesize surrounding views that resolve single-view ambiguity.","marker":"[21]"},{"why":"DreamGaussian baseline for single-image 3D generation that the symmetry-aware modifications build on.","marker":"[34]"},{"why":"NeRD, the prior state-of-the-art single-image symmetry detector and the formulation of reflection symmetry used throughout.","marker":"[48]"},{"why":"Objaverse, one of the two 3D repositories combined to build the large-scale training set (84,789 objects).","marker":"[6]"},{"why":"ShapeNet, the other 3D repository in the training set.","marker":"[5]"},{"why":"GSO, the real scanned-object dataset used for zero-shot evaluation of detection and generation.","marker":"[7]"},{"why":"OmniObject3D, the second real scanned-object dataset used for zero-shot evaluation.","marker":"[40]"}],"fun_headline_variants":["Zero-shot mirror plane detection boosts 3D generation","Single image mirror plane detection improves 3D generation","Symmetry from one photo enhances 3D generation","Detecting mirror planes in single images sharpens 3D models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The automatic ground-truth symmetry labels, produced by sampling candidate planes, reflecting point clouds, thresholding Chamfer distance, and refining with ICP, are accurate and perceptually aligned for both the training and evaluation objects.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot mirror plane detection boosts 3D generation","Single image mirror plane detection improves 3D generation","Symmetry from one photo enhances 3D generation","Detecting mirror planes in single images sharpens 3D models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000478,"raw_usage":{"total_tokens":2361,"prompt_tokens":933,"completion_tokens":1428,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1362}},"tokens_in":549,"tokens_out":1428,"duration_ms":10067,"temperature":1.0,"reasoning_tokens":1362,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:30:49.023106+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take symmetric CAD models with known, explicit mirror planes, render single images, and run both the automatic labeling protocol and Reflect3D on them. If the labeling protocol's planes disagree with the CAD-defined planes by more than a few degrees, or if the detector's error against the CAD-defined planes is much larger than the reported average geodesic distance (about $13^\\circ$ on GSO), then the benchmark labels — not the detector — are carrying the reported accuracy.","supporting_citations":[{"cited_title":"Dinov2: Learning robust visual features without supervision","cited_arxiv_id":null,"evidence_quote":"Frozen DINOv2 encoder that supplies the geometry-aware image features the symmetry decoder attends to."},{"cited_title":"Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation","cited_arxiv_id":null,"evidence_quote":"DreamGaussian baseline for single-image 3D generation that the symmetry-aware modifications build on."},{"cited_title":"NeRD: Neural 3d reflection symmetry detector","cited_arxiv_id":null,"evidence_quote":"NeRD, the prior state-of-the-art single-image symmetry detector and the formulation of reflection symmetry used throughout."},{"cited_title":"Objaverse: A universe of annotated 3d objects","cited_arxiv_id":null,"evidence_quote":"Objaverse, one of the two 3D repositories combined to build the large-scale training set (84,789 objects)."},{"cited_title":"Shapenet: An information-rich 3d model repository","cited_arxiv_id":null,"evidence_quote":"ShapeNet, the other 3D repository in the training set."},{"cited_title":"Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation","cited_arxiv_id":null,"evidence_quote":"OmniObject3D, the second real scanned-object dataset used for zero-shot evaluation."}],"review_version":1}