{"id":"c3cb77ca-133b-44dd-8abb-51b1ef05793f","arxiv_id":"2507.01009","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A VAE trained on normalized distance matrices of object contours yields shape descriptors invariant to similarity transforms and point re-indexing, and these descriptors improve downstream shape classification.","lead":"ShapeEmbed encodes the outline of an object as a distance matrix and trains a variational autoencoder to produce a shape descriptor that is invariant to position, size, rotation, reflection, and where along the outline you start. The method outperforms classical and learned baselines on shape classification of digits, MPEG-7 shapes, and biological images, which could help microscopy image analysis.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Indexation invariance is asserted as guaranteed but the architecture is admitted non-equivariant; nothing in the loss enforces latent equality, and no experiment measures it.","rationale":"The reader's weakest assumption identifies the same point: indexation invariance is not directly verified, the encoder is admitted to be not truly shift-equivariant, and the min-over-reindexation loss could mask non-invariance. My reading of the full text confirms this and sharpens it: even in the ideal case, Eq. 2 cannot enforce encoder invariance because the minimum over k,o is taken over ground-truth matrices, not over encoder outputs; it only makes the decoder tolerant to indexation. Thus the phrase 'guaranteed to map ... to the same latent vector' is stronger than anything the architecture or loss can deliver without an explicit invariance penalty or a direct test. The empirical evaluation is otherwise solid: distance-matrix representation gives translation/rotation invariance by construction, normalization gives scale invariance, ablations on sMNIST/sMPEG-7 and rMNIST/rMPEG-7 support the practical value of the components, and the biological applications (BBBC010, MEF) are sensible. The central problem is confined to the indexation-invariance guarantee, which is both the most novel claim and the least supported one. A direct latent-equality test, as proposed, would settle whether the concern lands; if the test passes, the paper's stated guarantee is vindicated empirically despite the acknowledged architectural non-equivariance. Because the issue is addressable with a relatively small experiment and the current evidence otherwise supports the method, conditional acceptance is the appropriate verdict: the condition is to run and report the re-indexation latent-equality test, and ideally to release code. I do not see grounds for rejection, as the classification results and ablations do not depend on the unverified guarantee, but the central claim should not stand without the test.","tokens_in":23290,"tokens_out":2234,"duration_ms":127904,"concrete_test":"Take a trained ShapeEmbed encoder and a fixed set of, say, 100 contours with N=64 points. For each contour, construct all 2N equivalent distance matrices D^{k,o} via Eq. 1, encode each to a latent vector, and compute the maximum pairwise Euclidean distance within each equivalence class. Report this 'intra-indexation latent diameter' normalized by the median distance between latents of different shapes. If the diameter is not near zero (e.g., comparable to inter-shape distances), the 'guaranteed to map ... to the same latent vector' claim fails as stated. To settle whether the min-loss masks encoder non-invariance, retrain ShapeEmbed with circular padding but replace Eq.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that equivalent re-indexings of an outline's distance matrix map to the same latent vector (Section 3.2, 'guaranteed to map all possible equivalent indexings of successive outline points to the same latent vector'). This claim rests on two linked assumptions that are both unverified. First, the authors themselves state that because ResNet-18 uses stride and pooling, 'the result is no longer truly shift equivariant or invariant' even with circular padding. So exact encoder-side invariance is not guaranteed by the architecture. Second, the min-over-reindexations loss (Eq. 2) supervises only the decoder: it selects the version of the ground-truth matrix closest to the reconstruction, and therefore cannot force the encoder to produce identical latents for D and D^{k,o}. The objective is minimized even if the encoder separates indexations and the decoder simply aligns to whichever indexation was preserved; the loss can absorb non-invariance. The experiments do not isolate this: the 'No index. inv.' ablation removes circular padding and the custom loss together, so any observed degradation could be caused by either component, and downstream logistic-regression accuracy is a coarse proxy that tolerates residual intra-class latent variation. No experiment reports latent-code distances across the 2N re-indexings of a single contour, so the load-bearing invariance assertion remains untested. This is not a disagreement with the empirical classification results, which may well be valid; it is a gap between the stated guarantee and the evidence provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ShapeEmbed, a self-supervised variational autoencoder that encodes ordered 2D contour points through normalized Euclidean distance matrices into a latent shape descriptor. The authors claim invariance to translation, scaling, rotation, reflection, and contour point re-indexing. The encoder is a ResNet-18 modified with circular padding, and the decoder is trained with a reconstruction loss that minimizes over all 2N equivalent re-indexings of the input distance matrix. The method is evaluated on MNIST, MPEG-7, BBBC010, and MEF datasets against classical descriptors, SimCLR, MAE variants, and O2VAE, with additional ablations and a robustness study on degraded segmentations.","tokens_in":23576,"tokens_out":3848,"duration_ms":48612,"significance":"If the invariance claims hold, ShapeEmbed would be a practically valuable contribution: it would provide the first VAE-based shape descriptor that is fully invariant to similarity transformations and to contour re-indexing, with demonstrated strengths in downstream classification and in generative exploration of bioimaging shape spaces. The empirical work is extensive and carefully reported: multiple baselines, multiple metrics with 5-fold cross-validation, separate ablations for rotation/translation and for regularization terms, a segmentation-noise robustness study, and qualitative latent-space analyses. The central limitation is that the paper's strongest advertised property, exact indexation invariance, is asserted but not directly verified, and the architecture's own text acknowledges that exact shift equivariance is lost in strided and pooling layers.","major_comments":[{"comment":"The sentence 'our VAE is thus guaranteed to map all possible equivalent indexings of successive outline points to the same latent vector' is stronger than what the architecture supports. The authors themselves state that, because ResNet-18 uses stride and pooling, 'the result is no longer truly shift equivariant or invariant' even with circular padding. Circular padding in convolutional layers does not restore exact equivariance through strided convolutions and pooling. No experiment in Section 4 measures latent vectors for the 2N re-indexed distance matrices of a single contour. This is a load-bearing assertion for contribution 1, so the manuscript should either provide a direct invariance test (for example, maximum pairwise latent distance or cosine similarity across D^{k,o} for held-out contours) or explicitly weaken the claim to approximate/practical invariance.","section":"Section 3.2"},{"comment":"The min-over-reindexations reconstruction loss supervises only the decoder: it selects the version of the ground-truth matrix closest to the decoder output. This loss is minimized even if the encoder preserves indexation information, because the decoder can align to whichever indexation was preserved. Nothing in the objective directly enforces equality of latent codes across the 2N equivalent matrices. The claim in Section 3.2 that the encoder 'is unable to distinguish' between re-indexings is therefore not implied by the loss alone. The authors should add an explicit invariance term or, at minimum, an empirical measurement showing that encoder outputs for D and D^{k,o} coincide.","section":"Equation (2)"},{"comment":"The 'No index. inv.' ablation removes circular padding and the custom indexation-invariant loss simultaneously, replacing both with constant padding and standard MSE reconstruction. Consequently, any observed drop in F1-score cannot be attributed to either component individually. Because this is the only experiment directly tied to indexation invariance, the ablation should separate the two changes. Moreover, downstream logistic-regression accuracy is a coarse proxy that tolerates residual intra-class latent variation; a direct latent-equality metric is needed to support the invariance claim.","section":"Table 2"}],"minor_comments":[{"comment":"The rows for MEF and BBBC010 appear to be swapped relative to Section 4.1: Section 4.1 reports 1,407 objects for BBBC010 and 26,198 objects for MEF, while Table 5 assigns 1,407 to MEF and 26,198 to BBBC010. The maximum/minimum outline sizes appear swapped as well.","section":"Table 5"},{"comment":"The code repository is given as 'https://github.com/link_to_be_added_in_camera-ready_version', which is a placeholder. Since reproducibility is a stated goal, a working repository link or an explicit statement of availability is needed.","section":"Section 4 / Supplementary B"},{"comment":"The 'ShapeEmbed+Sz' variant is used in Table 4 but not defined in the main text. Please specify whether the saved distance-matrix norm is concatenated to the latent vector before the logistic regression classifier and how that feature is normalized.","section":"Section 4.5"},{"comment":"The sentence 'ShapeEmbed performs well (F1-score≤ 0.7)' appears to be a typo: the reported values and Figure 12 indicate that the intended statement is F1-score ≥ 0.7.","section":"Supplementary Section G.3"},{"comment":"There are small typographical and citation issues, including 'not to to the same solution' in Section 3.4 and the spelling 'Lorense' for the marching-squares reference, which should be 'Lorensen'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is well founded and I agree with the conditional verdict. The empirical evaluation is strong and the method appears to work well in practice, but the central theoretical claim of guaranteed indexation invariance is not supported by the architecture or by the current experiments. The requested direct invariance test and the split of the indexation-invariance ablation are feasible within the manuscript's scope, so I do not see grounds for rejection. The placeholder code link should also be resolved before any final acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the method itself is sensible and the empirical work is genuinely good: distance-matrix inputs, circular padding, mirrored summation, and the min-over-reindexations loss make for a useful practical recipe, and the classification gains over strong baselines on MNIST, MPEG-7, BBBC010, and MEF are real. Second, the headline claim that ShapeEmbed is 'guaranteed' to map all re-indexed contours to the same latent vector is not supported by the evidence they present. That gap is fixable, but it should be fixed before publication.\n\nThe new content is the specific combination of ingredients: feeding normalized distance matrices into a VAE with a circular-padded ResNet encoder, summing mirrored inputs, and using the custom reconstruction loss. Prior work used distance matrices for shape comparison and VAEs for shape descriptors, but not this package, and the paper is honest about the lineage. The ablations are useful, particularly the scaling/indexation and rotation/translation tables. The latent-space visualizations and generative-sampling experiments are a nice bonus and strengthen the practical significance.\n\nThe soft spots are concentrated in the invariance claim. Section 3.2 admits that stride and pooling break exact shift equivariance, so the word 'guaranteed' is doing more work than the architecture can support. The min-over-reindexations loss (Eq. 2) supervises the decoder, not the encoder; it can be minimized even if the encoder separates indexations, as long as the decoder aligns to whichever indexation was preserved. And the 'No index. inv.' ablation removes circular padding and the custom loss together, so you cannot attribute the degradation to either component. What's missing is a direct measurement: for a single contour, do all 2N re-indexed distance matrices actually produce the same latent vector? That experiment is easy to run and would settle the matter. Also, the code is only available at a placeholder link, which hampers verification.\n\nNone of this undermines the empirical classification results, which probably hold up. The issue is a mismatch between the theoretical guarantee and the evidence. As it stands, this is a paper for bioimage-analysis researchers and anyone working on unsupervised shape descriptors. It deserves a serious referee, but the review should insist on an explicit indexation-invariance test and a public code release. If those land, I'd be happy to cite it and use it.","headline":"ShapeEmbed is a solid, well-evaluated method for learning shape descriptors from distance matrices, but the paper's central guarantee of point-indexation invariance is overstated and needs a direct test before I'd trust it.","tokens_in":24098,"tokens_out":1268,"would_cite":true,"duration_ms":15664,"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":"This paper claims that a self-supervised variational autoencoder trained on Euclidean distance matrices of 2D contours produces shape descriptors that are invariant to translation, scaling, rotation, reflection, and contour point…","keywords":["shape descriptors","variational autoencoder","distance matrix","geometric invariance","self-supervised learning","contour quantification","biological imaging","representation learning"],"falsifier":"Take a single contour, compute its distance matrix, generate all 2N re-indexed versions (N starting points times two travel directions), push each through the trained encoder, and measure the maximum pairwise distance between the resulting latent codes. If any two re-indexings yield codes that differ by more than the typical within-class spread of the latent space, the guaranteed-invariance claim is false, and the paper reports no such direct test.","tokens_in":23075,"feed_emoji":"📐","tokens_out":4429,"duration_ms":46366,"temperature":0.7,"pith_summary":"ShapeEmbed aims to show that a self-supervised variational autoencoder can turn the outline of any simply connected 2D object into a fixed-length descriptor that does not change when the object is translated, scaled, rotated, mirrored, or when the contour's starting point and travel direction change. The paper's evidence is a set of classification experiments in which a logistic regression on ShapeEmbed's latent codes beats classical shape statistics, Fourier descriptors, contrastive learning, masked autoencoders, and the orientation-invariant O2VAE on MNIST, MPEG-7, and two biological datasets. The practical payoff would be a shape representation that needs no pre-alignment or normalization of objects before analysis, which matters for microscopy where cells and organisms appear at arbitrary pose and size.","feed_headline":"Autoencoder yields shape codes immune to rotation, size, indexing","feed_subtitle":"Learning from pairwise point distances lets one latent vector cover all poses of a contour.","key_machinery":"The load-bearing object is the Euclidean distance matrix of the contour points: an N by N array of pairwise point distances that is translation- and rotation-invariant by construction and scale-invariant after division by its Frobenius norm. The argument is carried by three components: circular padding in every convolution and pooling layer of a ResNet-18 encoder to make it shift-equivariant under diagonal re-indexing; summing the encoder outputs of the matrix and its horizontal and vertical mirror to add reflection invariance; and a reconstruction loss that takes the minimum over all 2N equivalent indexations so the decoder is never forced to commit to one indexing.","core_discovery":"The central claim is that encoding the Euclidean distance matrix of a contour, rather than the segmentation mask, with a VAE whose encoder uses circular padding and whose decoder is trained with a min-over-reindexations loss yields latent vectors that are invariant to similarity transformations and to contour re-indexing. The paper argues that distance matrices are inherently invariant to translation and rotation, become scale-invariant after Frobenius normalization, and that the remaining indexation ambiguity, the 2N equivalent matrices from choice of origin and travel direction, is removed by the encoder design and loss, so the same latent code is produced for every equivalent indexing. It further reports that this descriptor outperforms the tested baselines in shape classification on natural image benchmarks and biological imaging datasets, and that adding object size back as an optional feature improves results on biology tasks.","pith_inferences":["If the invariance claim holds exactly, ShapeEmbed could act as a drop-in shape encoder for unsupervised phenotype discovery, where structure in the latent space is inspected directly rather than through a trained classifier.","The min-over-reindexations loss may let the decoder absorb residual encoder non-invariance rather than enforcing exact latent equality, so an explicit equality test would be needed to separate the two effects.","The consistent gains from adding size back suggest the most useful practical descriptor is a two-part code: one scale-invariant shape vector plus one scalar scale channel, rather than a single invariant vector.","The same distance-matrix-plus-indexation-invariance recipe could transfer to other self-supervised architectures than VAEs, provided their encoders can be made shift-equivariant under the 2N re-indexing group."],"forward_implications":["Shape descriptors can be extracted without alignment, centering, or size normalization of input objects, removing a preprocessing step that classical shape pipelines require.","The same trained model can reconstruct outlines from latent codes via multidimensional scaling, enabling visualization and generative sampling of shapes directly from the descriptor space.","Reintroducing object size as an extra feature alongside the invariant code improves classification on biological datasets, so size can be treated as an optional, task-dependent channel.","The method applies to simply connected 2D contours and would need new machinery for contours with holes or for genuine 3D surfaces.","In biological imaging, the unsupervised latent structure can reveal mismatches between experimental labels and actual phenotypes, as demonstrated on the C. elegans dataset."],"supporting_citations":[{"why":"Supplies the variational autoencoder framework and the KL divergence loss that ShapeEmbed builds on.","marker":"[Kingma and Welling, 2014]"},{"why":"Supplies the ResNet-18 backbone that is modified with circular padding for the encoder.","marker":"[He et al., 2016]"},{"why":"The direct competitor O2VAE that ShapeEmbed must beat and whose rotation-invariance approach it contrasts with its own.","marker":"[Burgess et al., 2024]"},{"why":"Provides the theory of Euclidean distance matrices that the input representation relies on.","marker":"[Dokmanic et al., 2015]"},{"why":"Supplies multidimensional scaling, the method used to reconstruct outlines from distance matrices.","marker":"[Cox and Cox, 2000]"},{"why":"Supplies marching squares, the algorithm used to extract simply connected pixel outlines from segmentation masks.","marker":"[Lorense, 1987]"},{"why":"The SimCLR contrastive learning baseline whose representations are compared against ShapeEmbed's.","marker":"[Chen et al., 2020]"},{"why":"The masked autoencoder baseline whose three vision transformer variants are compared against ShapeEmbed's.","marker":"[He et al., 2022]"}],"fun_headline_variants":["ShapeEmbed: one code for every contour pose and indexing","Distance-matrix learning yields shape codes invariant to similarity and indexing","Self-supervised contour embeddings immune to similarity and reindexing","One latent vector covers all similarity transforms and reindexings","Euclidean distance matrices give shape codes invariant to similarity and indexing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole invariance story depends on the claim that replacing padding with circular padding makes the ResNet-18 encoder shift-equivariant enough that all 2N re-indexed distance matrices of a contour map to the same latent code, even though the paper admits that stride and pooling break true shift equivariance.","fun_headline_variants_meta":{"raw":{"variants":["ShapeEmbed: one code for every contour pose and indexing","Distance-matrix learning yields shape codes invariant to similarity and indexing","Self-supervised contour embeddings immune to similarity and reindexing","One latent vector covers all similarity transforms and reindexings","Euclidean distance matrices give shape codes invariant to similarity and indexing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001719,"raw_usage":{"total_tokens":6750,"prompt_tokens":846,"completion_tokens":5904,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":5819}},"tokens_in":462,"tokens_out":5904,"duration_ms":59388,"temperature":1.0,"reasoning_tokens":5819,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:00:32.952906+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a single contour, compute its distance matrix, generate all 2N re-indexed versions (N starting points times two travel directions), push each through the trained encoder, and measure the maximum pairwise distance between the resulting latent codes. If any two re-indexings yield codes that differ by more than the typical within-class spread of the latent space, the guaranteed-invariance claim is false, and the paper reports no such direct test.","supporting_citations":[{"cited_title":"Euclidean distance matrices: essential theory, algorithms, and applications","cited_arxiv_id":null,"evidence_quote":"Provides the theory of Euclidean distance matrices that the input representation relies on."},{"cited_title":"Multidimensional scaling","cited_arxiv_id":null,"evidence_quote":"Supplies multidimensional scaling, the method used to reconstruct outlines from distance matrices."},{"cited_title":"A high resolution 3d surface construction algorithm","cited_arxiv_id":null,"evidence_quote":"Supplies marching squares, the algorithm used to extract simply connected pixel outlines from segmentation masks."},{"cited_title":"A simple framework for contrastive learning of visual representations","cited_arxiv_id":null,"evidence_quote":"The SimCLR contrastive learning baseline whose representations are compared against ShapeEmbed's."}],"review_version":1}