Pith. sign in

REVIEW 4 major objections 6 minor 24 references

The paper proposes that domain randomization for training insect detectors can be done entirely by perturbing the parameters of 3D Gaussian Splatting — shifting spherical-harmonic colors and injecting procedural noise — instead of extractin

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 04:12 UTC pith:7KJRMSSF

load-bearing objection Useful meshless DR pipeline idea, but the paper sells it as robust without running the experiment that would demonstrate it. the 4 major comments →

arxiv 2607.22890 v1 pith:7KJRMSSF submitted 2026-07-24 cs.GR cs.CVcs.LG

Meshless Domain Randomization via Explicit Parameter Perturbation of 3D Gaussian Splatting

Felipe Nunes Carbone de Carvalho , Joyce de Morais Souza , Alan de Aguiar , Charles Morphy D. Santos , Jo\~ao Paulo Gois This is my paper
classification cs.GR cs.CVcs.LG
keywords domain randomization3D Gaussian Splattingspherical harmonicsprocedural noisesynthetic datainsect identificationsim-to-realmeshless rendering
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes a meshless domain-randomization pipeline for generating synthetic training images of complex organic subjects, mainly insects. Instead of extracting polygonal meshes from 3D Gaussian Splatting reconstructions, it perturbs the radiance-field parameters directly: a photometric branch scales and shifts the base color and rotates first-order spherical-harmonic lighting, while a procedural branch replaces the base color with multi-scale 3D noise and removes view-dependent terms. The randomized fields are composited over stochastic backgrounds in a rasterization engine that also outputs bounding boxes automatically. The central claim is that this direct tensor manipulation expands the feature-space diversity of the synthetic data, overcoming the narrow texture and lighting constraints of the original captures while preserving the morphological identity of the specimens. Evidence is given through feature-embedding dispersion, but no detector is trained to confirm end-to-end sim-to-real benefit.

Core claim

The authors claim that domain randomization for complex organic subjects can be performed without any mesh-based modeling. They work directly on the parameter tensor of a 3D Gaussian Splatting reconstruction: for each Gaussian primitive they perturb the spherical-harmonic coefficients that encode base color and view-dependent lighting. The photometric branch applies a global scaling and offset to the base color and rotates the first-order lighting terms to simulate changing illumination; the procedural branch replaces the base color with multi-scale 3D spatial noise and zeros out all higher-order terms, so the rendered subject shows only its geometric shape. The modified fields are then comp

What carries the argument

The central object is the per-Gaussian parameter tensor of the splatting model, specifically the spherical-harmonic (SH) coefficient set attached to each primitive. The SH decomposition separates the view-independent base color (degree zero) from view-dependent reflection terms (higher degrees), and this split carries the argument: the photometric pipeline perturbs the base color and low-order lighting terms, while the procedural pipeline replaces the base color and eliminates all view-dependent components. A spatio-optical pruning step removes low-opacity, highly elongated splats that would otherwise amplify into artifacts under perturbation. The supporting mechanism is a rasterization engi

Load-bearing premise

The argument stands or falls on treating the measured increase in feature-space dispersion as evidence that the synthetic images are better training data; no detector is trained in this paper, so if the extra variance does not improve task-level robustness on real images, the central claim has no direct support.

What would settle it

Train the same object detector on (a) canonical unperturbed renders and (b) images from the full DR pipeline, fixing all other training settings, and evaluate both on held-out real video frames; if the DR-trained detector does not achieve higher mean average precision than the canonical-trained one, the core claim that these datasets are 'robust' is refuted. This is the experiment the paper lists as immediate future work.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Synthetic training sets for species identification become feasible for specimens that are semi-transparent, hairy, or otherwise impractical to mesh, with no manual texturing or surface extraction.
  • Annotations come for free from the rendering engine, so large labeled datasets can be generated automatically across arbitrary camera poses.
  • The generated images sample a wider feature distribution than the original captures, so classifiers trained on them should lean less on superficial texture patterns.
  • Any existing 3D Gaussian Splatting reconstruction can be fed through the same perturbation recipe, making the pipeline applicable beyond insects.
  • A detector trained on these images should transfer better to real photographs; demonstrating this is the paper's stated next step.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: the same SH-perturbation mechanism could be applied to any 3D Gaussian Splatting reconstruction—plants, medical anatomy, manufactured parts—wherever appearance randomization is useful but mesh extraction is lossy; the biggest payoff is for specimens with hair-like or semi-transparent structures.
  • Inference: because the procedural branch zeroes all view-dependent SH terms, its output is effectively a diffuse, relit version of the specimen; pairing the pipeline with an inverse-rendering step that separates albedo from lighting could allow physically consistent relighting and further increase diversity without breaking morphology.
  • Inference: feature-space variance is a proxy, not proof, of sim-to-real transfer; a stronger test would be to train a detector on the DR images and compare it against one trained on canonical renders, measuring mean average precision on real frames. The paper's own future-work list suggests this is the intended next step.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a meshless domain-randomization framework that operates directly on the parameter space of 3D Gaussian Splatting (3DGS), avoiding polygon-mesh extraction. The workflow consists of a spatio-optical pruning stage, two stochastic perturbation branches — a Photometric pipeline that modifies SH coefficients (DC color shift/scaling, first-order Wigner rotation, higher-band specular scaling) and a Procedural pipeline that replaces DC appearance with 3D spatial noise and zeroes higher-order bands — and a Unity-based rasterizer that composites the perturbed radiance fields over random backgrounds and extracts 2D bounding boxes. The method is evaluated on two insect specimens (moth and fly) by embedding real frames, canonical 3DGS renderings, and DR renderings with a pre-trained ResNet-50, projecting them with UMAP, and reporting silhouette scores. The paper claims that the increased silhouette separation demonstrates that direct tensor manipulation broadens feature-space coverage and 'provides a meshless alternative for generating robust datasets.'

Significance. If the result were fully validated, the contribution would be useful and timely: it would provide a low-cost, meshless alternative to traditional DR for complex organic specimens, with explicit SH-perturbation equations, automatic annotations, and a path toward scalable synthetic-data generation for biodiversity applications. The perturbative equations in §3.2–§3.3 are clear, simple, and internally consistent, and the paper genuinely avoids the mesh-extraction bottleneck. However, the current evidence supports only the weaker statement that DR renderings occupy a more dispersed region of a pretrained feature space. The central claim of generating 'robust datasets' — i.e., data that improve sim-to-real generalization — is not directly tested, and the paper itself defers detector training to future work. The separate contributions of photometric DR, procedural DR, and background compositing are not ablated, and no statistical or task-level validation is provided. The idea is promising, but the evidence is preliminary.

major comments (4)
  1. [§4.1 and §5] The headline claim that direct tensor manipulation 'provides a meshless alternative for generating robust datasets' is not established by the evidence. The only quantitative support is the silhouette score between real and synthetic ResNet-50 feature clusters, which is interpreted as confirmation that the DR pipeline 'expands feature-space variance' and 'provides the varied training data necessary to encourage sim-to-real generalization' (§4.1). A higher silhouette score measures increased separation between the real and synthetic clusters, not task-relevant variation within the synthetic cluster; it may indicate a larger domain gap rather than beneficial augmentation. No detector or classifier is trained, and §5 explicitly defers YOLO evaluation to future work. The paper should either add a downstream detection/classification benchmark or temper the conclusion to 'feature-space variance
  2. [§4.1] The quantitative evaluation is a point estimate from 100 images per category and a single reconstruction per specimen. No error bars, confidence intervals, repeats over random seeds, or statistical tests are reported, so the observed silhouette differences (moth 0.1436→0.2235; fly 0.1532→0.1847) may be within run-to-run variation. Moreover, no ablation separates the contributions of the photometric pipeline, the procedural pipeline, and the background compositing. Because the randomized-background DR category includes stochastic backgrounds, the measured dispersion could be driven almost entirely by background variation rather than by SH or noise perturbations. To support the central claim, the authors should provide per-module ablations and repeatability statistics.
  3. [§4.1, Eq. (4)] It is unspecified whether the silhouette score is computed on the 2048-dimensional ResNet-50 activations or on the 2D UMAP projection. The text says features are 'projected into 2D using UMAP' and then refers to 'high-dimensional feature space.' If the score uses the raw features, state that explicitly; if it uses the UMAP embedding, justify why a stochastic 2D projection is used for a quantitative metric, report UMAP hyperparameters (n_neighbors, min_dist, metric), and show that the results are stable across multiple UMAP runs. As written, the numerical results are not reproducible.
  4. [§2 and §4] The motivation is that mesh-based DR suffers from lossy surface extraction and texture-bias problems, but no comparison with any mesh-based alternative is presented. Showing that DR renderings are more dispersed than canonical 3DGS renderings does not demonstrate an advantage over, e.g., SuGaR mesh extraction followed by a conventional mesh-based DR pipeline. A direct comparison — on morphology preservation, downstream detection accuracy, or annotation cost — is needed to substantiate the stated contribution of a 'meshless alternative.'
minor comments (6)
  1. [§3.1] Specify whether the opacity threshold 'below 10%' refers to alpha values in [0,1] multiplied by 100 or to percentile ranks. Also clarify how the 'longest principal axis exceeding 10 times the shortest' is computed from the covariance represented by scale s and quaternion q.
  2. [§3.2] The term 'AC' (for higher-order SH components) is used without definition; define it at first use. In Eq. (2), c^ℓ is overloaded: for ℓ=1 it is a 3-vector, while for ℓ>1 it is a (2ℓ+1)-vector. Use explicit per-degree indexing to avoid ambiguity.
  3. [§3.3] The 'multi-scale 3D spatial noise' N(μ_i) is described only verbally. Please specify the noise type (e.g., Perlin, value, fractal), the scales or octaves, and how Φ stochastically selects between the step function and identity mapping. These details are necessary for reproducibility.
  4. [§3.4 / §4.1] The method of aligning canonical renderings to the original camera trajectories is not described. The statement that canonical control images are 'rendered from manually aligned camera poses that closely matched original trajectories' should be supported by the alignment procedure or by quantitative pose-error estimates.
  5. [§4.1] The segmentation of 'real-world video frames with backgrounds removed' is not specified. State whether segmentation was manual, mask-based, or performed by an off-the-shelf model, and note the potential impact of segmentation artifacts on the ResNet-50 features.
  6. [General] Minor typographical inconsistencies: 'SuGAR' appears as 'SuGaR' in the references; 'E FA-GS' should be 'EFA-GS'; 'A C Wigner' appears as a section heading and should be typeset consistently. The UMAP caption in Fig. 4 may rely on color that is not distinguishable in grayscale; add distinct marker shapes.

Circularity Check

0 steps flagged

No significant circularity found; the central evaluation gap is external validity, not a self-referential derivation.

full rationale

The paper's derivation chain is not circular. The perturbation equations (Eqs. 1-3) directly manipulate 3DGS SH coefficients and procedural noise; they do not encode or presuppose the ResNet-50/UMAP/silhouette evaluation outcome. The evaluation uses a pretrained external feature extractor, and the reported silhouette increase (0.1436/0.1532 to 0.2235/0.1847) is a measured empirical quantity, not a parameter fitted to the evaluation. The perturbation ranges are disclosed as empirical hyperparameters, not fitted to make the silhouette score come out in a particular direction. There are no load-bearing self-citations: none of the cited references appear to be by the present authors, and no uniqueness theorem or prior result is invoked to forbid alternatives. The paper does make an unsupported leap when it equates feature-space dispersion with task-level sim-to-real robustness, and it explicitly defers detector training to future work (Section 5: 'our immediate priority is to evaluate the proposed synthesized datasets by training state-of-the-art taxonomic object detectors (e.g., YOLO variants)'). That is a missing validation, not circularity: the claim is under-supported rather than equivalent to its inputs. The acknowledged limitations and future-work statements are consistent with an honest, incomplete evaluation, not with a circular derivation.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or representational entities; it perturbs existing 3DGS parameters. However, the method depends on many hand-chosen or unspecified sampling distributions and on feature-space proxies for task relevance. The absence of code/data and the manual preprocessing steps further increase the implicit assumptions.

free parameters (7)
  • Pruning thresholds (opacity, elongation) = opacity < 0.10; longest axis > 10x shortest
    Empirical hyperparameters chosen to remove artifacts before perturbation (Section 3.1).
  • Δ_color shift bound = [-1.2, 1.2]^3
    Empirical hyperparameter for simulating white-balance shifts (Eq. 1, Section 3.2).
  • λ exposure scalar = unspecified distribution
    Stochastic scalar factor in Eq. 1; its sampling distribution is not described.
  • γ_spec log-normal scaling = unspecified log-normal parameters
    Stochastic specular scaling in Eq. 2; the log-normal parameters are not reported.
  • Wigner rotation R angle sampling = unspecified
    Rotation applied to l=1 SH band in Eq. 2; how rotation angles are drawn is not specified.
  • Procedural noise N(µ_i) = unspecified multi-scale 3D noise
    Noise type, scales, and octaves are not specified (Eq. 3, Section 3.3).
  • Φ threshold / identity and C_amp / C_shift mapping = unspecified
    Binarization threshold and color mapping gains are 'dynamically scaled' but not defined (Eq. 3).
axioms (5)
  • domain assumption 3DGS reconstructions from SuGaR/EFA-GS with default parameters are faithful enough for domain randomization
    Reconstructions are manually cropped but no reconstruction-quality metric is reported (Section 4).
  • domain assumption Unity rasterizer faithfully renders the perturbed 3DGS parameter tensors
    No comparison between the Unity renderer and a reference 3DGS rasterizer is provided (Section 3.4).
  • domain assumption ResNet-50 embeddings are a valid perceptual/morphological similarity measure for insects
    Used for UMAP and silhouette evaluation without validation for insect morphology (Section 4.1).
  • domain assumption Silhouette-score increase implies useful domain randomization while preserving identity
    The interpretation that dispersion equals robustness is asserted, not tested with a downstream task (Section 4.1–5).
  • domain assumption Restricting explicit rotation to l=1 SH and scaling higher orders captures needed lighting variation
    Justified by prior avatar work [14,15] and computational cost, but not independently validated on insects (Section 3.2, limitations).

pith-pipeline@v1.3.0-alltime-deepseek · 7884 in / 10317 out tokens · 106149 ms · 2026-08-01T04:12:00.807599+00:00 · methodology

0 comments
read the original abstract

Domain Randomization (DR) is a standard technique for closing the Sim-to-Real gap, yet traditional DR pipelines rely on classical computer graphics rendering driven by polygon meshes. For complex organic subjects, such as insect specimens, extracting and rendering textured meshes is challenging. To address this issue, we propose a meshless DR framework that operates on the parameter space of 3D Gaussian Splatting (3DGS). Our method employs two independent perturbation pipelines to synthesize randomized training datasets. First, a Photometric DR pipeline alters the baked illumination and color balance by modulating the Spherical Harmonics (SH) coefficients. Second, a Procedural DR pipeline isolates the subject's geometric shape by replacing its original textures with 3D spatial noise. Finally, these perturbed radiance fields are composited over stochastically varied backgrounds using a rasterization engine. Our parameter manipulation provides a meshless alternative for generating robust datasets for complex geometries.

Figures

Figures reproduced from arXiv: 2607.22890 by Alan de Aguiar, Charles Morphy D. Santos, Felipe Nunes Carbone de Carvalho, Jo\~ao Paulo Gois, Joyce de Morais Souza.

Figure 1
Figure 1. Figure 1: Baseline 3DGS reconstructions. Top: Representative free-hand multi-view video [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Photometric DR pipeline outputs applied to the moth and fly 3DGS models. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Procedural DR pipeline outputs. Native textures are overridden by 3D spatial noise [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: 2D UMAP feature-space projections of pre-trained ResNet-50 embeddings across four [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 4 canonical work pages

  1. [1]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimkuehler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Trans. Graph., vol. 42, no. 4, Jul. 2023. [Online]. Available: https://doi.org/10.1145/3592433 8

    doi:10.1145/3592433 2023

  2. [2]

    Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness

    R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel, “Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness.” inInternational Conference on Learning Representations, 2019. [Online]. Available: https://openreview.net/forum?id=Bygh9j09KX

    2019

  3. [3]

    Accurate detection and identification of insects from camera trap images with deep learning,

    K. Bjerge, J. Alison, M. Dyrmann, C. E. Frigaard, H. M. R. Mann, and T. T. Høye, “Accurate detection and identification of insects from camera trap images with deep learning,”PLOS Sustainability and Transformation, vol. 2, no. 3, pp. 1–18, 03 2023. [Online]. Available: https://doi.org/10.1371/journal.pstr.0000051

    work page doi:10.1371/journal.pstr.0000051 2023

  4. [4]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” in 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE Press, 2017, p. 23–30. [Online]. Available: https://doi.org/10.1109/IROS.2017.8202133

    arXiv 2017

  5. [5]

    Leveraging synthetic data produced from museum specimens to train adaptable species classification models,

    J. D. Blair, K. Khidas, and K. E. Marshall, “Leveraging synthetic data produced from museum specimens to train adaptable species classification models,”PLOS ONE, vol. 20, no. 9, pp. 1–15, 09 2025. [Online]. Available: https://doi.org/10.1371/journal.pone.0329482

    work page doi:10.1371/journal.pone.0329482 2025

  6. [6]

    Classification of seeds using domain randomization on self-supervised learning frameworks,

    V. Margapuri and M. Neilsen, “Classification of seeds using domain randomization on self-supervised learning frameworks,” in2021 IEEE Symposium Series on Computational Intelligence (SSCI), 2021, pp. 01–08. [Online]. Available: https: //doi.org/10.1109/SSCI50451.2021.9659998

    arXiv 2021

  7. [7]

    Synthetic dataset generation for autonomous mobile robots using 3d gaussian splatting for vision training,

    A. Deogan, W. Beks, P. Teurlings, K. De Vos, M. Van Den Brand, and R. Van De Molengraft, “Synthetic dataset generation for autonomous mobile robots using 3d gaussian splatting for vision training,” in2025 IEEE 21st International Conference on Automation Science and Engineering (CASE), 2025, pp. 2802–2807. [Online]. Available: https://doi.org/10.1109/CASE5...

    arXiv 2025

  8. [8]

    Automatic generation of synthetic data and object detection datasets in virtual environments based on tomato-harvesting robot vision,

    T. Ushiroji and T. Fujinaga, “Automatic generation of synthetic data and object detection datasets in virtual environments based on tomato-harvesting robot vision,” Smart Agricultural Technology, vol. 14, p. 101947, 2026. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2772375526001711

    2026

  9. [9]

    Automated specimen triage for dark taxa: Deep learning enables orientation, sex identification and anatomical segmentation from robotic imaging,

    H. Shirali, L. W¨ uhrl, L. Lee, N. Klug, R. Meier, C. Pylatiuk, and E. Hartop, “Automated specimen triage for dark taxa: Deep learning enables orientation, sex identification and anatomical segmentation from robotic imaging,”Systematic Entomology, vol. 51, no. 1, p. e70039, 2026. [Online]. Available: https://resjournals.onlinelibrary.wiley.com/doi/abs/10....

    2026

  10. [10]

    From blender to farm: Transforming controlled environment agriculture with synthetic data and swinunet for precision crop monitoring,

    K. Aghamohammadesmaeilketabforoosh, J. Parfitt, S. Nikan, and J. M. Pearce, “From blender to farm: Transforming controlled environment agriculture with synthetic data and swinunet for precision crop monitoring,”PLOS ONE, vol. 20, no. 4, pp. 1–17, 04

  11. [11]

    A survey on 3d gaussian splatting applications: Segmentation, editing, and generation,

    S. He, P. Ji, Y. Yang, C. Wang, J. Ji, Y. Wang, and H. Ding, “A survey on 3d gaussian splatting applications: Segmentation, editing, and generation,” 2026. [Online]. Available: https://arxiv.org/abs/2508.09977

    Pith/arXiv arXiv 2026

  12. [12]

    A survey on 3d gaussian splatting,

    G. Chen and W. Wang, “A survey on 3d gaussian splatting,”ACM Comput. Surv., vol. 58, no. 12, May 2026. [Online]. Available: https://doi.org/10.1145/3807511

    doi:10.1145/3807511 2026

  13. [15]

    On the skinning of gaussian avatars,

    N. Zioulis, N. Kotarelas, G. Albanis, S. Thermos, and A. Chatzitofis, “On the skinning of gaussian avatars,” 2025. [Online]. Available: https://arxiv.org/abs/2509.11411

    arXiv 2025

  14. [16]

    Gs-ir: 3d gaussian splatting for inverse rendering,

    Z. Liang, Q. Zhang, Y. Feng, Y. Shan, and K. Jia, “Gs-ir: 3d gaussian splatting for inverse rendering,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 21 644–21 653. [Online]. Available: https://doi.org/10.1109/CVPR52733.2024.02045

    arXiv 2024

  15. [17]

    Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing,

    J. Gao, C. Gu, Y. Lin, Z. Li, H. Zhu, X. Cao, L. Zhang, and Y. Yao, “Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing,” inComputer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds. Cham: Springer Nature Switzerland, 2025, pp. 73–89. [Online]. Available: htt...

    doi:10.1007/978-3-031-72995-9 2024

  16. [18]

    3dgaa: Realistic and robust 3d gaussian-based adversarial attack for autonomous driving,

    Y. Zhang, L. Wang, J. Zhao, W. Zhao, F. Zhou, Y. Dang, and J. Yin, “3dgaa: Realistic and robust 3d gaussian-based adversarial attack for autonomous driving,” 2025. [Online]. Available: https://arxiv.org/abs/2507.09993

    arXiv 2025

  17. [19]

    Unity gaussian splatting,

    A. Pranckeviˇ cius, “Unity gaussian splatting,” 2023, accessed: 2026-03-06. [Online]. Available: https://github.com/aras-p/UnityGaussianSplatting

    2023

  18. [20]

    Low-frequency first: Eliminating floating artifacts in 3d gaussian splatting,

    J. Wang, P. Zhou, C. Li, R. Quan, and J. Qin, “Low-frequency first: Eliminating floating artifacts in 3d gaussian splatting,” 2025. [Online]. Available: https: //arxiv.org/abs/2508.02493

    arXiv 2025

  19. [21]

    [Online]

    SuperSplat, 2026. [Online]. Available: https://superspl.at/editor

    2026

  20. [22]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778. [Online]. Available: https://doi.org/10.1109/CVPR.2016.90

    doi:10.1109/cvpr.2016.90 2016

  21. [23]

    Umap: Uniform manifold approximation and projection for dimension reduction,

    L. McInnes, J. Healy, and J. Melville, “Umap: Uniform manifold approximation and projection for dimension reduction,”arXiv preprint arXiv:1802.03426, 2018. [Online]. Available: https://arxiv.org/abs/1802.03426

    Pith/arXiv arXiv 2018

  22. [24]

    Silhouettes: A graphical aid to the interpretation and validation of cluster analysis,

    P. J. Rousseeuw, “Silhouettes: A graphical aid to the interpretation and validation of cluster analysis,”Journal of Computational and Applied Mathematics, vol. 20, pp. 53–65, 1987. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ 0377042787901257

    1987

  23. [25]

    Rapid, without focus stacking, 3d photogrammetric digitization of cockroaches,

    J. Berger, B. Bignon, R. Cornette, F. Legendre, and A. Delapr´ e, “Rapid, without focus stacking, 3d photogrammetric digitization of cockroaches,”PLOS ONE, vol. 20, no. 12, pp. 1–16, 12 2025. [Online]. Available: https://doi.org/10.1371/journal.pone.0336893 10

    work page doi:10.1371/journal.pone.0336893 2025

  24. [2025]

    Available: https://doi.org/10.1371/journal.pone.0322189

    [Online]. Available: https://doi.org/10.1371/journal.pone.0322189

    work page doi:10.1371/journal.pone.0322189