Pith. sign in

REVIEW 3 major objections 5 minor 59 references

3DMPE: 3D Multi-Perspective Embedding

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

Pith's one-line read 3DMPE recovers 3D point clouds from partial 2D projections by optimizing a visibility-masked multi-perspective stress, without category-specific training.

desk verdict Clean, training-free geometric recovery after matching: visibility-masked MPSE stress works under known correspondences; scope is the real limit, not the math. read the letter →

arxiv 2607.04898 v1 pith:4OPKEWOB submitted 2026-07-06 cs.CV cs.CG

classification cs.CVcs.CG
keywords 3Dpointcloudreconstructionmulti-viewgeometrymulti-perspectiveembeddingincompletepairwisedistancesvisibility-maskedstresstraining-freeStructure-from-Motion
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

When several incomplete 2D views of the same object are available, each showing only a subset of the points, recovering a single consistent 3D configuration is hard because the usual distance matrices are full of missing entries. This paper shows that the missingness can be handled directly: a multi-perspective stress function simply multiplies each term by the product of the two points’ visibility flags, so only observed pairs contribute. Minimizing that objective with stochastic gradient descent and a smart MDS-based start recovers the 3D coordinates; when the camera angles themselves are unknown they can be estimated jointly. Because the method never trains on object categories, it works for any geometry once correspondences and visibility are known. Experiments on ShapeNet and Pix3D confirm that a few well-spread views with modest noise still produce low Chamfer and Earth-Mover distances, making the approach a practical geometric module after feature matching has already been done.

What carries the argument

Visibility-masked multi-perspective stress (Eq. 2): the ordinary multi-perspective MDS stress multiplied by the product of the two points’ visibility indicators, so incomplete pairwise distances are simply ignored rather than imputed.

What would settle it

On a controlled ShapeNet mesh whose true correspondences are known, replace them with the sparse, noisy tracks produced by a standard structure-from-motion pipeline on the same rendered views; if the resulting Chamfer distance rises above the paper’s own acceptability threshold of roughly 0.2 while the same optimizer with ground-truth matches stays well below it, the claim that the method works once “geometric observations are available” fails for realistic correspondence quality.

Watch

Extended reading notes

Core claim

Given two or more partially observed 2D projections of an unknown 3D point cloud, together with known cross-view correspondences and per-view visibility, a single non-convex stress objective that zeros every term involving a missing point recovers a consistent 3D configuration (and, when needed, the projection maps themselves) without any category-specific training.

Load-bearing premise

Cross-view point correspondences and visibility labels must already be supplied as reliable geometric input; the optimizer has no independent signal if those matches are sparse or systematically wrong.

Editorial extensions

If this is right

  • A training-free geometric stage can sit after any correspondence engine and still produce usable 3D point clouds from a handful of partial views.
  • Four or five viewpoints that each hide some points are usually enough once every point appears in at least three views.
  • The same objective can be used whether camera angles are known or must be estimated jointly.
  • Local errors in distances or correspondences degrade reconstruction only gradually, so modest noise does not destroy the solution.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Modern dense matchers that already output visibility or scores could feed 3DMPE almost unchanged, turning it into a drop-in geometric back-end.
  • Because the method never sees category labels, it should transfer immediately to non-ShapeNet domains (medical landmarks, industrial parts, archaeological fragments) once correspondences exist.
  • The smart MDS initialization that averages incomplete distance matrices may itself be a useful warm-start for other multi-view optimization problems.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies 3D point-cloud reconstruction from two or more partially observed 2D projections when cross-view correspondences and per-view visibility are known. It introduces 3DMPE, a training-free extension of Multi-Perspective Simultaneous Embedding that minimizes a visibility-masked multi-perspective stress (Eq. 2) over 3D coordinates and, in the variable-projection setting, over projection maps. Fixed- and variable-projection variants are optimized by minibatch SGD with an MDS-based smart initialization. Experiments on ShapeNet and Pix3D report CD, EMD, and a custom ROA metric under changes in initialization, number of views, visibility, point-cloud size, angle span, additive distance noise, and local correspondence errors, and include an exploratory COLMAP integration.

Significance. Under the stated premise that correspondences and visibility are already available, the work supplies a clean, category-agnostic geometric recovery module that complements both classical SfM pipelines and learning-based image-to-shape methods. Strengths include an explicit incomplete-distance objective, systematic ablations (initialization, views, visibility, noise), public code, and an honest scope statement that sparse or unreliable matches break the method (as the COLMAP trial illustrates). If the empirical findings hold, 3DMPE is a useful building block for multi-view pipelines that already produce tracks, rather than a replacement for end-to-end reconstruction.

major comments (3)
  1. Tables 3 and 5 (and the qualitative Tables 1 and 4) place 3DMPE CD/EMD numbers beside PSGN and 3D-LMNet. Those methods take raw images and different supervision; 3DMPE takes geometric observations with known correspondences. The paper notes complementarity in the introduction and §4, yet the side-by-side tables still invite direct ranking. Either remove the quantitative learning-method columns or reframe them strictly as context with an explicit “not comparable inputs” caveat in the table captions and text, so the central geometric-recovery claim is not overstated.
  2. Because 3DMPE is positioned as the geometric recovery stage after correspondences (Introduction; §5), the primary experimental baseline should be classical multi-view geometry for partial observations (triangulation with visibility, robust bundle adjustment, or COLMAP’s sparse reconstruction given the same tracks), not only the MDS average-distance baseline and image-based networks. The COLMAP experiment in §4.4 tests correspondence extraction, not reconstruction quality under identical geometric inputs; without that comparison the claim that 3DMPE “effectively reconstructs” relative to standard geometric tools remains incompletely supported.
  3. Algorithm 2 and §3.2 leave the variable-projection parameterization underspecified: Q is updated with a projection operator Π, but the manifold (orthogonal/Stiefel constraints, gauge freedom, initialization of viewpoints) is not stated in closed form. The variable-projection setting is a main contribution; without a precise description (or a pointer to the released code’s exact parameterization), independent reimplementation and verification of the joint estimation results are difficult.
minor comments (5)
  1. Notation drifts between P (Eqs. 1–4, Algorithm 1) and Q (Algorithm 2) for projection parameters; unify or define the relationship explicitly.
  2. Fig. 4 axis labels contain OCR-style corruption (“Dis ance Ma rix”, “P oint Correspondence”); regenerate for readability.
  3. The ROA metric is useful; state in the main text (not only the appendix) whether the reported numbers use the iterative or SVD alignment, and whether they are scale-normalized consistently with CD/EMD.
  4. Runtime plots (Figs. 5, 12) would benefit from a brief note on hardware and whether the implementation is single-threaded NumPy/SciPy or uses GPU acceleration, given the ~200 s figures for 2048 points.
  5. Minor typos: “3-Dimensional” vs “3D” in the title/abstract consistency; “fixed”/“varying” vs “variable” projection wording; “pyoint” package name in §4.5.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: 3DMPE is a training-free optimization of a visibility-masked multi-perspective stress, evaluated against external ground-truth meshes with independent metrics.

full rationale

The paper formulates reconstruction as minimization of the visibility-masked stress S3DREC (Eq. 2), an explicit extension of the MPSE stress (Eq. 1) that zeros missing pairwise distances via the product of visibility indicators. This is a standard non-convex geometric optimization problem (fixed or variable projections), solved by SGD with optional MDS-based smart initialization; the baseline MDS average and the optimized 3DMPE results are reported separately, so the claim is not equated with the initializer. Reconstruction quality is measured by Chamfer Distance, Earth Mover Distance, and ROA against held-out ShapeNet/Pix3D ground-truth point clouds after alignment, not by restating the stress value. Self-citation of MPSE supplies the parent multi-perspective embedding method but does not force the numerical reconstruction results or uniqueness of the recovered clouds. No parameter is fitted to a subset of the evaluation data and then re-labeled a prediction; noise and visibility ablations are controlled experiments against external truth. The correspondence/visibility assumption is an explicit scope boundary (Introduction, §5), not a circular reduction of the geometric claim. Therefore the derivation chain is self-contained and non-circular.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The central claim rests on classical MDS/MPSE stress minimization plus the modeling choice that 2D observations are pairwise Euclidean distances under linear projections with known matches and binary visibility. Free knobs are optimizer and experiment controls, not physics constants. No new physical entities; the invented pieces are the masked objective, the ROA alignment metric, and the smart-init pipeline.

free parameters (4)
  • SGD iteration budget / early-stop thresholds = T≤300; tol=1e-4
    Default max 300 iterations; stop when objective or gradient norm < 10^{-4} (§4). Directly affects reported reconstructions.
  • Stochastic sampling constant c
    Probability of including distance entries in minibatch ξ ~ Ξ(D,c) (Algorithms 1–2); controls stochastic gradient noise.
  • Adaptive learning-rate scheme μ
    Barzilai–Borwein-style adaptive steps (Eq. 6) with separate μ_X, μ_Q in variable-projection mode; hand-chosen initial rates.
  • Viewpoint angle range θ_r and visibility count per point = default θ_r=360°; often 3 of 5 views
    Synthetic view generation and occlusion model (θ_s, θ_e; points visible in k of K views) are experimental design choices that strongly affect CD (Figs. 7–10).
assumptions (5)
  • domain assumption Cross-view point correspondences are known a priori for all landmarks used in reconstruction.
    Stated as problem input in Abstract/Introduction; primary limitation in §5.
  • domain assumption Each view provides Euclidean pairwise distances among currently visible points under a linear projection P^{(k)}.
    Stress (Eq. 2) compares D^{(k)}_{ij} to ||P^{(k)} x_i − P^{(k)} x_j||; synthetic data from rotation matrices (Eq. 12) and ray tracing.
  • domain assumption Visibility is binary and known; missing points zero entire pairwise contributions via α_i α_j.
    Definition of S_3DREC (Eq. 2) and visibility vectors in §3.
  • ad hoc to paper Non-convex stress minimization via SGD with MDS-based initialization yields high-quality local minima in practice.
    Authors note non-convexity (§3.2) and rely on smart init + early stopping rather than global optimality guarantees.
  • standard math Standard MDS/MPSE stress and Euclidean embedding geometry.
    Inherited from Shepard MDS and MPSE stress (Eq. 1) cited in §2.1.
invented entities (2)
  • 3DMPE visibility-masked multi-perspective stress S_3DREC
    purpose: Objective that allows incomplete multi-view distance matrices for 3D reconstruction.
    Eq. (2) is the paper’s central formal object; independent evidence is empirical reconstruction quality, not external theory.
  • ROA (RMSE-Optimize-Align) metric
    purpose: Correspondence-aware alignment error after optimal rigid transform (Eq. 11 / SVD appendix).
    Custom evaluation metric introduced in §3.3; not a physical entity but a paper-specific score.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3DMPE: 3D Multi-Perspective Embedding." pith.science (2026). https://pith.science/paper/4OPKEWOB

@misc{pith2026260704898,
  author       = {Pith},
  title        = {Pith review of: 3DMPE: 3D Multi-Perspective Embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4OPKEWOB}},
  note         = {Machine review of arXiv:2607.04898}
}
read the original abstract

We study 3D point cloud reconstruction from multiple partially observed 2D projections. Given two or more projections of an unknown 3D point cloud, together with cross-view point correspondences and visibility information, our goal is to recover a consistent 3D configuration when different views contain different subsets of points. We propose 3D Multi-Perspective Embedding (3DMPE), an optimization-based, training-free method that reconstructs the 3D point cloud and, in the variable-projection setting, jointly estimates the projection maps. 3DMPE extends Multi-Perspective Simultaneous Embedding to accommodate missing points and incomplete pairwise distance information across views. We consider both fixed-projection and variable-projection settings. Unlike learning-based reconstruction methods that infer shape from raw images and often depend on training data, 3DMPE operates on geometric observations with established correspondences and does not require category-specific training. Experiments on ShapeNet and Pix3D evaluate reconstruction quality using Chamfer Distance, Earth Mover Distance, and RMSE-Optimize-Align (ROA), and examine the effects of initialization, the number of views, point visibility, and several noise regimes, including noisy distances and erroneous correspondences. The results demonstrate that 3DMPE can effectively reconstruct point clouds from partial multi-view geometric observations.

Figures

Figures reproduced from arXiv: 2607.04898 by the authors.

Figure 1
Figure 1. Full pipeline of 3DMPE. Given multiple 2D snapshots and the visibility vectors (1 if a point is present in a view and 0 otherwise), in this example 3, but the algorithm works for arbitrary number of inputs, we first compute the pairwise distance matrices for each snapshot the algorithm reconstructs the 3D point cloud. In practice, 3D reconstruction is often addressed using Structure-from-Motion (SfM) pipelines [3, 1… view at source ↗
Figure 2
Figure 2. Pipeline for ROA metric. Here 0.64 is the ROA value for current iteration of gradient descent update loop. controlled comparison is with the proposed MDS-based baseline. We addition￾ally include published results for learning-based methods and an exploratory comparison with COLMAP for context; these methods use different inputs and assumptions and are therefore not directly comparable to 3DMPE. 4.1 Datasets and Expe… view at source ↗
Figure 3
Figure 3. Effect of initialization: blue corresponds to random and orange to smart ini￾tialization. The Effect of Smart Initialization As the objective functions of 3DMPE with fixed (3) and varying projections (4) are non-convex, initialization is im￾portant for the SGD step. In this section we analyze the effect of smart initializa￾tion (see Sec. 3.3) on the performance of 3DMPE. We run 3DMPE with smart initialization and co… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Noise robustness analysis of 3DMPE (1024 points, 5 viewpoints). x-axis shows the percentage of corrupted points. Top curves: baseline model (squares); bottom curves: 3DMPE (circles). corrupted points results in a slight increase in CD values while staying in the desire…
Figure 5
Figure 5. Figure 5: Runtime (seconds) of 3DMPE with the ShapeNet dataset: (a-b) show the im￾pact of changing the number of perspectives (using 512 points) and (c-d) show the impact of changing the number of points (with 4 projections). We note that 3DMPE with variable projections solves a…
Figure 6
Figure 6. Figure 6: EMD and CD values with varying number of points on ShapeNet (4 viewpoints, each point visible from 3+ viewpoints). We analyze the sensitivity of 3DMPE to the percentage of hidden points as follows: fix a parameter ϑ, which represents the number of points that we want t…
Figure 7
Figure 7. Figure 7: 3DMPE metrics on ShapeNet for varying viewpoints (512 points; all visible for 23 views, otherwise each point appears in one fewer view). 1     [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Metric analysis of 3DMPE on ShapeNet dataset for points visible in a number of viewpoints. All experiments are done for 512 points with 8 viewpoints. cept 2 and 3 where all points are visible). The experiments show that 3DMPE reconstructs the 3D point cloud properly if…
Figure 9
Figure 9. Figure 9: Points visible in at least 1 . . . 5 viewpoints for ray tracing with five equidistant views. 3DMPE reconstructions for the corresponding points are shown above each bar; reconstruction metrics are reported in the table. 3     3      …
Figure 10
Figure 10. Figure 10: Performance of 3DMPE with different constraints on viewpoint positions. The x-axis (logarithmic scale) denotes the maximum angle distance between each viewpoint from the center of point cloud [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: and [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Runtime (seconds) of 3DMPE on Pix3D dataset: (a-b) show the impact of changing the number of perspectives (using 512 points) and (c-d) show the impact of changing the number of points (with 4 projections). 1  1 1  [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: EMD, CD and ROA metrics on a varying number of points on Pix3D dataset. Each experiment is done with 4 viewpoints and each point is visible from at least 3 viewpoints [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Maximum correspondences per perspective. With 703 perspectives and 65,373 feature points, the low correspondence count leads to poor reconstruction despite the noise robustness of 3DMPE. Although the two pipelines use substantially different inputs and assump￾tions, t…
Figure 15
Figure 15. Figure 15: ROA Metric analysis of 3DMPE on ShapeNet dataset for (a-b) points visible in a number of viewpoints and (c-d) varying number of viewpoints. All experiments are done for 512 points with 8 viewpoints. For (c-d) each point is visible 1 less than the number of viewpoints …
Figure 16
Figure 16. Figure 16: ROA metrics on a varying number of points on ShapeNet dataset (similar to figure 6a and 6b). Each experiment is done with 4 viewpoints and each point is visible from at least 3 viewpoints [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 2 canonical work pages

  1. [1]

    Ultramicroscopy 234, 113460 (2022) 22 Vahan Huroyan, Md Rahat-uz-Zaman, and Stephen Kobourov

    Altingövde, O., Mishchuk, A., Ganeeva, G., Oveisi, E., Hebert, C., Fua, P.: 3d reconstruction of curvilinear structures with stereo matching deep convolutional neural networks. Ultramicroscopy 234, 113460 (2022) 22 Vahan Huroyan, Md Rahat-uz-Zaman, and Stephen Kobourov

  2. [2]

    Berger, M., Tagliasacchi, A., Seversky, L.M., Alliez, P., Guennebaud, G., Levine, J.A., Sharf, A., Silva, C.T.: A survey of surface reconstruction from point clouds. Comput. Graph. Forum 36(1), 301–329 (2017)

  3. [3]

    Bianco, S., Ciocca, G., Marelli, D.: Evaluating the performance of structure from motion pipelines. J. Imaging 4(8), 98 (2018)

  4. [4]

    CoRR abs/1512.03012 (2015)

    Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. CoRR abs/1512.03012 (2015)

  5. [5]

    In: The 2006 IEEE International Joint Conference on Neural Network Pro- ceedings

    Cheng, W.C.: Neural-network-based photometric stereo for 3d surface reconstruc- tion. In: The 2006 IEEE International Joint Conference on Neural Network Pro- ceedings. pp. 404–410 (2006). https://doi.org/10.1109/IJCNN.2006.246710

  6. [6]

    In: Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII

    Choy, C.B., Xu, D., Gwak, J., Chen, K., Savarese, S.: 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In: Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII. Lecture Notes in Computer Science, vol. 9912, pp. 628–644. Springer (2016)

  7. [7]

    Blender Foun- dation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender

    Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foun- dation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender. org

  8. [8]

    In: submitted to the proceedings of the GraVisMa workshop, Plzen

    Deul, C., Burger, M., Hildenbrand, D., Koch, A.: Raytracing point clouds using geometric algebra. In: submitted to the proceedings of the GraVisMa workshop, Plzen. Citeseer (2010)

Show all 59 references
  1. [9]

    In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017

    Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object re- construction from a single image. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 2463–2471. IEEE Computer Society (2017)

  2. [10]

    Com- mun

    Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Com- mun. ACM 24(6), 381395 (jun 1981). https://doi.org/10.1145/358669.358692, https://doi.org/10.1145/358669.358692

  3. [11]

    In: Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI

    Girdhar, R., Fouhey, D.F., Rodriguez, M., Gupta, A.: Learning a predictable and generative vector representation for objects. In: Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI. Lecture Notes in Com...

  4. [12]

    Artificial intelligence (AI) 14, 16 (2021)

    Hamid, M.S., Manap, N., Hamzah, R.A., Kadmin, A.F.: Stereo matching algorithm based on hybrid convolutional neural network and directional intensity difference. Artificial intelligence (AI) 14, 16 (2021)

  5. [13]

    Kybernetika 46(5), 926–937 (2010)

    Häming, K., Peters, G.: The structure-from-motion reconstruction pipeline - a survey with focus on short image sequences. Kybernetika 46(5), 926–937 (2010)

  6. [14]

    IEEE Trans

    Han, X., Laga, H., Bennamoun, M.: Image-based 3d object reconstruction: State- of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell. 43(5), 1578–1604 (2021)

  7. [15]

    Harltey, A., Zisserman, A.: Multiple view geometry in computer vision (2. ed.). Cambridge University Press (2006)

  8. [16]

    IEEE Trans

    Hossain, M.I., Huroyan, V., Kobourov, S.G., Navarrete, R.: Multi-perspective, simultaneous embedding. IEEE Trans. Vis. Comput. Graph. 27(2), 1569–1579 (2021)

  9. [17]

    Insafutdinov, E., Dosovitskiy, A.: Unsupervised learning of shape and pose with differentiable point clouds. In: Advances in Neural Information Processing Systems 3DMPE: 3D Multi-Perspective Embedding 23 31: Annual Conference on Neural Information Processing Systems 2018, NeurI...

  10. [18]

    In: Advances in Neural Information Processing Systems 30: Annual Conference on Neural Infor- mation Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA

    Kar, A., Häne, C., Malik, J.: Learning a multi-view stereo machine. In: Advances in Neural Information Processing Systems 30: Annual Conference on Neural Infor- mation Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA. pp. 365–376 (2017)

  11. [19]

    In: 30th British Machine Vision Conference, BMVC 2019, Cardiff, UK, September 9-12, 2019

    Klokov, R., Verbeek, J., Boyer, E.: Probabilistic reconstruction networks for 3d shape inference from a single image. In: 30th British Machine Vision Conference, BMVC 2019, Cardiff, UK, September 9-12, 2019. p. 165. BMV A Press (2019)

  12. [20]

    In: 30th British Machine Vision Conference 2019, BMVC 2019, Cardiff, UK, September 9-12, 2019

    Kulon, D., Wang, H., Güler, R.A., Bronstein, M.M., Zafeiriou, S.: Single image 3d hand reconstruction with mesh convolutions. In: 30th British Machine Vision Conference 2019, BMVC 2019, Cardiff, UK, September 9-12, 2019. p. 45. BMV A Press (2019)

  13. [21]

    In: Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII

    Lei, J., Sridhar, S., Guerrero, P., Sung, M., Mitra, N.J., Guibas, L.J.: Pix2surf: Learning parametric 3d surface models of objects from images. In: Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII. Lecture Notes ...

  14. [22]

    In: Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, Septem- ber 29-October 4, 2024, Proceedings, Part LXXII

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, Septem- ber 29-October 4, 2024, Proceedings, Part LXXII. Lecture Notes in Computer Science, vol. 15130, pp. 71–91. Springer (2024)

  15. [23]

    Lin, C., Kong, C., Lucey, S.: Learning efficient point cloud generation for dense 3d object reconstruction. In: Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8t...

  16. [24]

    7114–7121

    pp. 7114–7121. AAAI Press (2018)

  17. [25]

    In: Proceed- ings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999

    Lowe, D.G.: Object recognition from local scale-invariant features. In: Proceed- ings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999. pp. 1150–1157. IEEE Computer Society (1999)

  18. [26]

    In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event

    Malitsky, Y., Mishchenko, K.: Adaptive gradient descent without descent. In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event. Proceedings of Machine Learning Research, vol. 119, pp. 6702–6712. PMLR (2020)

  19. [27]

    In: British Machine Vision Conference 2018, BMVC 2018, Newcastle, UK, September 3-6, 2018

    Mandikal, P., L., N.K., Agarwal, M., Radhakrishnan, V.B.: 3d-lmnet: Latent em- bedding matching for accurate and diverse 3d point cloud reconstruction from a single image. In: British Machine Vision Conference 2018, BMVC 2018, Newcastle, UK, September 3-6, 2018. p. 55. BMV A P...

  20. [28]

    In: IEEE Winter Conference on Applications of Computer Vision, W ACV 2019, Waikoloa Village, HI, USA, January 7-11, 2019

    Mandikal, P., Radhakrishnan, V.B.: Dense 3d point cloud reconstruction using a deep pyramid network. In: IEEE Winter Conference on Applications of Computer Vision, W ACV 2019, Waikoloa Village, HI, USA, January 7-11, 2019. pp. 1052–

  21. [29]

    In: 31st British Machine Vision Conference, BMVC 2020, Virtual Event, UK, September 7-10, 2020

    Michalkiewicz, M., Belilovsky, E., Baktashmotlagh, M., Eriksson, A.P.: A simple and scalable shape representation for 3d reconstruction. In: 31st British Machine Vision Conference, BMVC 2020, Virtual Event, UK, September 7-10, 2020. BMV A Press (2020)

  22. [30]

    IEEE Transactions on Intelligent Transportation Systems 21(4), 1514– 1524 (2020)

    Miclea, V.C., Nedevschi, S.: Real-time semantic segmentation-based stereo recon- struction. IEEE Transactions on Intelligent Transportation Systems 21(4), 1514– 1524 (2020). https://doi.org/10.1109/TITS.2019.2913883 24 Vahan Huroyan, Md Rahat-uz-Zaman, and Stephen Kobourov

  23. [31]

    In: Proceedings 2001 ICRA

    Neubert, J., Hammond, T., Guse, N., Do, Y., Hu, Y., Ferrier, N.: Automatic train- ing of a neural net for active stereo 3d reconstruction. In: Proceedings 2001 ICRA. IEEE International Conference on Robotics and Automation (Cat. No.01CH37164) (2001)

  24. [32]

    Acta Numer

    Özyesil, O., Voroninski, V., Basri, R., Singer, A.: A survey of structure from motion. Acta Numer. 26, 305–364 (2017)

  25. [33]

    The Annals of Math- ematical Statistics 22(3), 400 – 407 (1951)

    Robbins, H., Monro, S.: A Stochastic Approximation Method. The Annals of Math- ematical Statistics 22(3), 400 – 407 (1951)

  26. [34]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

    Rocco, I., Arandjelovic, R., Sivic, J.: Convolutional neural network architecture for geometric matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

  27. [35]

    Rubner, Y., Tomasi, C., Guibas, L.J.: The earth mover’s distance as a metric for image retrieval. Int. J. Comput. Vis. 40(2), 99–121 (2000)

  28. [36]

    github.io/datasets.html

    Schoenberger, J.L.: 4 datasets in colmap documentation (2023), https://colmap. github.io/datasets.html

  29. [37]

    In: CVPR

    Schönberger, J.L., Frahm, J.: Structure-from-motion revisited. In: CVPR. pp. 4104–4113 (2016)

  30. [38]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

    Seki, A., Pollefeys, M.: Sgm-nets: Semi-global matching with neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

  31. [39]

    Psychometrika 27(2), 125–140 (1962)

    Shepard, R.N.: The analysis of proximities: multidimensional scaling with an un- known distance function. Psychometrika 27(2), 125–140 (1962)

  32. [40]

    A K Peters (2000)

    Shirley, P.: Realistic ray tracing. A K Peters (2000)

  33. [41]

    In: A vidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T

    Shrestha, R., Hu, S., Gou, M., Liu, Z., Tan, P.: A real world dataset for multi- view 3d reconstruction. In: A vidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel A viv, Israel, October 23-27, 202...

  34. [42]

    In: Advances in Neural Information Processing Systems 32, Vancouver, BC, Canada

    Sridhar, S., Rempe, D., Valentin, J., Bouaziz, S., Guibas, L.J.: Multiview aggre- gation for learning category-specific shape reconstruction. In: Advances in Neural Information Processing Systems 32, Vancouver, BC, Canada. pp. 2348–2359 (2019)

  35. [43]

    SIAM Rev

    Stewart, G.W.: On the early history of the singular value decomposition. SIAM Rev. 35(4), 551–566 (1993)

  36. [44]

    In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018

    Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Freeman, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. ...

  37. [45]

    In: IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017

    Tatarchenko, M., Dosovitskiy, A., Brox, T.: Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In: IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017. pp. 2107–2115. IEEE Computer Society (2017)

  38. [46]

    In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017

    Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single- view reconstruction via differentiable ray consistency. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 209–217. IEEE Comp...

  39. [47]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15,

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotný, D.: VGGT: visual geometry grounded transformer. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15,

  40. [48]

    5294–5306

    pp. 5294–5306. Computer Vision Foundation / IEEE (2025) 3DMPE: 3D Multi-Perspective Embedding 25

  41. [49]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, W A, USA, June 16-22, 2024

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, W A, USA, June 16-22, 2024. pp. 20697–20709. IEEE (2024)

  42. [50]

    https://mathworld

    Weisstein, E.W.: Triangle point picking. https://mathworld. wolfram. com/ (1999)

  43. [51]

    In: Advances in Neural Information Processing Systems 29: Annual Conference on Neural Infor- mation Processing Systems, Barcelona, Spain

    Wu, J., Zhang, C., Xue, T., Freeman, B., Tenenbaum, J.: Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In: Advances in Neural Information Processing Systems 29: Annual Conference on Neural Infor- mation Processing Systems, Barcel...

  44. [52]

    In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015

    Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015. pp. 1912–1920. IEEE Computer Society (2015)

  45. [53]

    Constructive Approximation 26(2), 289–315 (2007)

    Yao, Y., Rosasco, L., Caponnetto, A.: On early stopping in gradient descent learn- ing. Constructive Approximation 26(2), 289–315 (2007)

  46. [54]

    In: 32nd British Machine Vision Conference 2021, BMVC 2021, Online, November 22-25, 2021

    Yu, Z., Yang, L., Chen, S., Yao, A.: Local and global point cloud reconstruction for 3d hand pose estimation. In: 32nd British Machine Vision Conference 2021, BMVC 2021, Online, November 22-25, 2021. p. 388. BMV A Press (2021)

  47. [55]

    CoRR abs/1510.05970 (2015)

    Zbontar, J., LeCun, Y.: Stereo matching by training a convolutional neural network to compare image patches. CoRR abs/1510.05970 (2015)

  48. [56]

    IEEE Trans

    Zhang, J., Yao, Y., Deng, B.: Fast and robust iterative closest point. IEEE Trans. Pattern Anal. Mach. Intell. 44(7), 3450–3466 (2022)

  49. [57]

    IEEE Trans

    Zheng, J.X., Pawar, S., Goodman, D.F.M.: Graph drawing by stochastic gradient descent. IEEE Trans. Vis. Comput. Graph. 25(9), 2738–2748 (2019)

  50. [58]

    IEEE Transactions on Neural Networks and Learning Systems (2019)

    Zhu, H., Jiao, L., Ma, W., Liu, F., Zhao, W.: A novel neural network for remote sensing image matching. IEEE Transactions on Neural Networks and Learning Systems (2019)

  51. [59]

    Mathematics 10(4), 542 (2022) 26 Vahan Huroyan, Md Rahat-uz-Zaman, and Stephen Kobourov Appendix Alternative Computation of the ROA Metric In Sec

    Zhu, Z., Iglesias, A., Zhou, L., You, L., Zhang, J.: Pde-based 3d surface reconstruc- tion from multi-view 2d images. Mathematics 10(4), 542 (2022) 26 Vahan Huroyan, Md Rahat-uz-Zaman, and Stephen Kobourov Appendix Alternative Computation of the ROA Metric In Sec. 3.3, we prop...

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.