Pith. sign in

REVIEW 4 major objections 5 minor 54 references

SurfR: Surface Reconstruction with Multi-scale Attention

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SurfR claims a single learned model can reconstruct surfaces from point clouds faster than existing methods while keeping near-top accuracy, by delaying query-point use until after multi-scale features are built.

desk verdict Solid architecture paper whose headline speed claim is undercut by its own timing table; deserves review after honest revision. read the letter →

arxiv 2506.08635 v1 pith:CMNX7PM2 submitted 2025-06-10 cs.CV

Siddhant Ranade , Gonçalo Dias Pais , Ross Tyler Whitaker , Jacinto C. Nascimento , Pedro Miraldo , Srikumar Ramalingam This is my paper
classification cs.CV
keywords surfacereconstructionimplicitneuralrepresentationsigneddistancefunctionpointcloudsmulti-scalefeaturescross-scaleattentionlazyquerysampling3Ddeeplearning
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

The paper sets out to show that a single learned model can reconstruct 3D surfaces from unorganized, noisy point clouds faster than each of the main baselines running at its own preferred resolution, while losing only a little reconstruction fidelity. The speed comes from separating the point cloud from the query points: SurfR first builds per-cell features at three grid scales in parallel, then samples those features at query points and fuses the scales with attention. This lazy-query design lets one point-cloud encoding answer many signed-distance queries cheaply. If the claim holds, interactive applications such as robotic grasping, AR/VR meshing, and SLAM get a practical path to real-time surface reconstruction without per-object fitting.

What carries the argument

The load-bearing mechanism is a parallel multi-scale cell-grid feature encoder with lazy query sampling. Space is partitioned into $s^3$ cells at scales $s \in \{1,4,16\}$; points are transformed to cell coordinates, encoded per point, max-pooled per cell, and then, only after all scales are built, sampled at the query points using inverse-distance-weighted nearest-neighbor interpolation. For each query, a single-head transformer encoder layer attends across the three scales, and the concatenated attended features are regressed to a sign logit and a magnitude by a five-layer MLP. Because the point-cloud encoding is computed once and reused for every query point, the cost of SDF evaluation decouples from the size of the feature-extraction network and scales with the number of query evaluations.

What would settle it

Run the paper's code and each baseline's code on the same GPU, measuring wall-clock time from a raw point cloud to a meshed SDF at every method's stated optimal resolution, repeated several times; if SurfR is not faster than P2S, POCO, and NKSR under identical conditions, the speed claim fails. A second check isolates the accuracy gap: on the max-noise ABC split, disable the box-filter sign propagation and remeasure Chamfer distance to see whether the documented bump artifacts explain the difference.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that a query-point-independent feature extraction is sufficient for high-fidelity implicit reconstruction. SurfR partitions the unit cube into $s^3$ cells for $s\in\{1,4,16\}$, encodes each cell with a local PointNet-style MLP and channel-wise max pooling, and later gathers per-query features by inverse-distance-weighted nearest-neighbor interpolation of local features and cell features. A single transformer attention layer then reweights the three scales for each query, and an MLP head regresses sign logits and magnitudes of the signed distance. In the reported experiments this reaches a Chamfer distance of $2.1$ on the no-noise ABC test set at resolution 256, against $1.8$ for P2S and $1.7$ for POCO, while taking about $7$ seconds per mesh compared with $232$ seconds for P2S and $51$ seconds for POCO. The paper also states its own limitations: fast sign propagation can leave bump-like artifacts, and stronger noise yields rougher surfaces.

Load-bearing premise

The headline speed advantage assumes that the inference times in the paper's comparison table were measured under comparable hardware, software implementations, and optimization effort, but the paper reports no inference GPU, no repeated timings, and no timing protocol, so the central speed claim stands or falls on that comparability.

Editorial extensions

If this is right

  • At 256 resolution, SurfR reports roughly $33\times$ and $7.3\times$ speedups over P2S and POCO, which would put full grid-based SDF extraction inside interactive editing and SLAM loop budgets.
  • The model generalizes to unseen ABC shapes, FAMOUS meshes, Thingi10K objects, and real-world structure-from-motion scans without any per-object training.
  • Because point-cloud encoding is query-independent, adding more query points or raising the output resolution grows the cost with the query count rather than with the input size, so one encoding can serve many evaluations.
  • The reported accuracy gap to P2S and POCO appears mainly in the max-noise setting, so the speed gain comes with a regime-dependent fidelity trade-off.

Reading between the lines

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

  • A natural extension the paper does not test is to precompute the cell features once and evaluate the SDF at multiple increasing resolutions from the same encoding, which the lazy-query design already permits and would further amortize the remaining cost.
  • The cell-based parallel structure suggests a streaming variant in which only cells containing newly arrived points are re-encoded, which would suit incremental SLAM or continuously updating sensor feeds.
  • The documented bump artifacts point to replacing the fixed box-filter sign propagation with a learned or confidence-weighted sign aggregation as a concrete way to close the remaining accuracy gap.
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

4 major / 5 minor

Summary. The paper proposes SurfR, an implicit surface reconstruction method for unorganized point clouds. It combines three components: lazy query feature sampling that defers the query point until after per-cell features are computed, a parallel multi-scale grid feature extractor based on PointNet-like encoders, and a cross-scale self-attention layer that fuses query features from different scales. An SDF head predicts sign logits and magnitude, which are combined into a signed distance, and the surface is extracted with marching cubes after a sign-propagation step. Experiments compare SurfR with six learning-based baselines (CON, P2S, SAP, IF-Net, POCO, NKSR) on ABC, FAMOUS, and Thingi10K datasets, reporting Chamfer distance, normal consistency, and inference time. The central claim, stated in the abstract and Section 4.4, is that SurfR is 'faster than all the baselines at their optimum resolution, with only a marginal loss in performance compared to the state-of-the-art.'

Significance. If the reported results are reliable, SurfR would be a practical contribution to real-time surface reconstruction: the lazy-query design and parallel multi-scale attention are plausible mechanisms for decoupling feature extraction from query evaluation, and the ablation study (Table 1) shows consistent improvements from each added component. The paper also provides an unusually broad comparison across six baselines and three datasets. However, the headline speed/accuracy trade-off claim is currently not established: it depends on a selective choice of SurfR's resolution, the speed numbers are reported without any measurement protocol, and the accuracy comparisons lack error bars or significance testing. The strengths of the paper are its clear architectural narrative and its extensive experimental coverage, but the central claim needs to be either substantiated with additional evidence or substantially weakened.

major comments (4)
  1. [Abstract and Section 4.4, Table 4] The abstract claims that SurfR is 'faster than all the baselines at their optimum resolution, with only a marginal loss in performance compared to the state-of-the-art.' This is only true for the SurfR@64 configuration: Table 4 lists NKSR at 0.5 s, while SurfR@128 takes 1.8 s and SurfR@256 takes 7 s. At SurfR@64, the average Chamfer distance in Table 3 is 2.4, compared to 1.9 for both P2S and POCO, a 26% relative degradation, and in the max-noise row the gap is 3.3 versus 2.7. The paper does not define what 'marginal' means, and it does not provide error bars or per-shape paired comparisons. The claim as stated is therefore not supported by the data; it must be either restricted to a clearly specified resolution with a quantitative threshold for 'marginal', or reformulated to acknowledge that at 128/256 surfR is not faster than NKSR.
  2. [Section 4.2 and Section 4.4, Table 4] The timing comparison lacks a measurement protocol. Table 4 reports inference times in seconds without specifying the hardware, the number of repetitions, whether the timings include the sign-propagation and marching-cubes steps, or whether the baselines were run with their own post-processing. Without this information, the speed figures are not verifiable and the claim that SurfR is the fastest at any configuration is not reproducible. The authors should provide the hardware details, timing methodology, and a breakdown of which pipeline components are included.
  3. [Section 4.4, Table 3] The quantitative comparisons are reported as single averages over the test sets, with no standard deviations, per-shape distributions, or statistical significance tests. Several of the key comparisons (e.g., SurfR@128 average Chamfer 2.1 vs. P2S 1.9 and POCO 1.9) involve differences that could easily be within run-to-run or model-seed variation. This is load-bearing because the paper uses these numbers to argue that the accuracy loss is 'marginal' and that SurfR is 'competitive with state-of-the-art.' I request that the authors add variance estimates or paired significance tests, at least for the main ABC and Thingi10K comparisons.
  4. [Section 4.1 and Section 4.4] The baseline comparison is not fully controlled. CON and IF-Net are trained on ShapeNet, while P2S, POCO, NKSR, and SurfR are trained on ABC; NKSR additionally receives normals estimated from the input point cloud, and the text states that the evaluation 'does not use additional refinement steps such as POCO and NKSR.' These differences could systematically affect both the accuracy and the timing comparisons. The authors should clarify how each baseline was obtained (pretrained weights or retrained), whether all baselines are evaluated with or without their native post-processing, and quantify the impact of the normal-estimation step on NKSR's results. As it stands, Table 3 mixes several confounds that the paper does not control for.
minor comments (5)
  1. [References and Table 3/4] The reference [17] is cited as 'P2S' in Tables 3 and 4, and in the baselines list. Reference [17] is Point2Mesh by Hanocka et al., which is not the P2S method. The Points2Surf method is reference [11]. This misassignment should be corrected.
  2. [Figure 7 caption] The caption contains grammatical errors and unclear phrasing: '(a) shows Bump-like surfaces show due to the sign propagation during evaluation' should be rephrased, and the figure does not clearly separate the two limitations described in the text.
  3. [Section 4.3, Table 1] The abbreviation 'LOD' is used without definition; it appears to mean 'level of detail,' but this should be explicitly stated.
  4. [Section 4.3] The ablations in Table 1 were performed with batch size 4 and only 100 training epochs, while the final model uses batch size 16 and 750 epochs. This difference could change the relative improvements of the ablations, and the authors should either match the training budget or discuss why the conclusions are expected to transfer.
  5. [General] The paper would benefit from a short 'Reproducibility' statement indicating whether the code and trained models will be released, as this is now common practice for benchmarking papers in this area.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SurfR's claims are empirical comparisons on held-out benchmarks, and design choices are validated by ablations rather than fitted to the target result.

full rationale

The paper's contributions are presented as empirical architecture choices and evaluated through controlled ablations and held-out test sets. The multi-scale scales {1,4,16}, the loss weights, and the sign-propagation threshold are selected via validation on ABC and then tested on unseen ABC, FAMOUS, and Thingi10K data; no model parameter is fitted to the headline Chamfer or timing numbers and then renamed a prediction. The accuracy and speed claims in Tables 3 and 4 are direct comparisons against external baselines on data not used for training, which is standard evidence rather than a circular derivation. The only self-citation, reference [13] (an author-prior work on an oriented-grid encoder), appears in a related-work catalog of object-specific methods and is not load-bearing for any of SurfR's design decisions or results. Concerns that the speed advantage rests on a favorable resolution choice (SurfR at 64) and that the timing protocol is unspecified are evidence-quality issues, not circularity: they do not show that any claim is equivalent to its own input by construction. Hence no circular step is present and the score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

SurfR is an empirical architecture paper. The central claims rest on standard ML components, public datasets, and a handful of hand-set hyperparameters; there are no new theoretical entities. The main unresolvable gaps are the unreported K, the absent inference timing protocol, and the under-trained ablations.

free parameters (6)
  • Multi-scale grid sizes = {1, 4, 16}
    Selected from Table 2 ablations; controls local versus global context and runtime.
  • Per-point feature dimension F1 = 64
    Hand-set in Section 4.2; capacity affects reconstruction detail and speed.
  • Per-cell feature dimension F2 = 128
    Hand-set in Section 4.2; capacity affects reconstruction detail and speed.
  • Loss weights lambda_mag, lambda_sgn, lambda_reg = 5.0, 2.0, 1e-6
    Hand-set in Section 4.2; balance magnitude, sign, and regularization.
  • Box filter size eps and sign propagation threshold t_update = 5, 1/3
    Hand-set in Section 3.7; directly trades reconstruction time against surface quality.
  • Number of nearest neighbors K in query feature sampler = not stated
    K is used in Equations 3 and 4 but its value is never reported; it directly controls the query feature content.
assumptions (5)
  • domain assumption Ground-truth signed distances computed from meshes are available and accurate for training.
    Section 4.1 uses mesh SDF values as supervision; if mesh normals or SDF computation are inaccurate, the learned SDF inherits the error.
  • domain assumption BlenSor-simulated time-of-flight scans are representative of real sensor noise.
    All quantitative tests use BlenSor-generated point clouds; only two real objects are shown qualitatively. If the simulation does not match deployment sensors, generalization claims weaken.
  • domain assumption The sign-propagation heuristic recovers the correct surface from truncated SDF signs.
    Section 3.7 repeatedly applies a box filter with threshold t_update=1/3; the paper's own limitations section notes this can create bump-like artifacts, so the heuristic is known to be imperfect.
  • standard math Marching cubes correctly extracts isosurfaces from the SDF grid.
    Section 3.7 invokes marching cubes [28]; this is a standard algorithm whose correctness is not in question.
  • domain assumption Transformer attention layers generalize as cross-scale feature fusion without overfitting.
    Section 3.4 uses one transformer layer; the paper validates it only on ABC at a reduced training budget, and its robustness to distribution shift is not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SurfR: Surface Reconstruction with Multi-scale Attention." pith.science (2026). https://pith.science/paper/CMNX7PM2

@misc{pith2026250608635,
  author       = {Pith},
  title        = {Pith review of: SurfR: Surface Reconstruction with Multi-scale Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMNX7PM2}},
  note         = {Machine review of arXiv:2506.08635}
}
read the original abstract

We propose a fast and accurate surface reconstruction algorithm for unorganized point clouds using an implicit representation. Recent learning methods are either single-object representations with small neural models that allow for high surface details but require per-object training or generalized representations that require larger models and generalize to newer shapes but lack details, and inference is slow. We propose a new implicit representation for general 3D shapes that is faster than all the baselines at their optimum resolution, with only a marginal loss in performance compared to the state-of-the-art. We achieve the best accuracy-speed trade-off using three key contributions. Many implicit methods extract features from the point cloud to classify whether a query point is inside or outside the object. First, to speed up the reconstruction, we show that this feature extraction does not need to use the query point at an early stage (lazy query). Second, we use a parallel multi-scale grid representation to develop robust features for different noise levels and input resolutions. Finally, we show that attention across scales can provide improved reconstruction results.

Figures

Figures reproduced from arXiv: 2506.08635 by the authors.

Figure 1
Figure 1. SurfR result and error/time trade-off: At the top, we show a sparse input point cloud and its reconstruction using the proposed method SurfR. At the bottom, we show our method and the baselines error vs. time trade-off comparison, with different resolutions (number below). We propose SurfR, a data-driven and efficient implicit sur￾face reconstruction method that produces high-fidelity mod￾els, striking the right bal… view at source ↗
Figure 2
Figure 2. Overview The input consists of point cloud P and query points Q. For each scale s ∈ {s0, s1, . . . , sS}, an encoder net￾work is used to extract per-cell features F c s . These features are sampled at the query points to get query point features F q s with our novel query-feature sampling technique. A self-attention mecha￾nism is applied to these features to obtain a new set of per-query features F a s , which are s… view at source ↗
Figure 3
Figure 3. Parallel Multi-scale feature extraction The input points P in (i) are transformed to cell coordinates Ps, at each scale s (ii), and each individual cell is processed in parallel as shown in the inset. Then, the points are encoded by PointNet e l s to get the per-point features F l s (iii). Further they are encoded by the cell en￾coder e c s and pooled cell-wise to get the per-cell features F c s (iv). (i) (ii) (iii)… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Query Feature Sampling The input to this stage con￾sists of Ps, F l s, F c s from the multi-scale feature extraction, and the query points Q. The query points ii are transformed to cell coordinates Qs (ii - top). The output per-query feature F q s (vi) is a concatenati…
Figure 5
Figure 5. Figure 5: Qualitative Baseline Comparison: the results of the proposed method are qualitatively on par with the baselines, or even better, particularly on the Famous and Thingi10K datasets, with different noise/density levels. SurfR preserves more detail even in sparsely sampled…
Figure 6
Figure 6. Figure 6: Reconstruction of Real-world Objects: SurfR generalizes well, even in noise and variable sampling density. The proposed method shows it can handle real-world objects taken from SfM methods [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Limitations: (a) shows Bump-like surfaces show due to the sign propagation during evaluation, and (b) Surface roughness in noisy input point cloud. appear when noise levels increase, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 50 canonical work pages

  1. [1]

    Neural Volume Super-Resolution

    Yuval Bahat, Yuxuan Zhang, Hendrik Sommerhoff, Andreas Kolb, and Felix Heide. Neural volume super-resolution. arXiv preprint arXiv:2212.04666, 2022. 3

    work page Pith review arXiv 2022

  2. [2]

    Poco: Point convo- lution for surface reconstruction

    Alexandre Boulch and Renaud Marlet. Poco: Point convo- lution for surface reconstruction. InIEEE/CVF Conf. Com- puter Vision and Pattern Recognition (CVPR), pages 6302– 6314, 2022. 2, 6, 8

    work page 2022

  3. [3]

    J. C. Carr, R. K. Beatson, J. B. Cherrie, T. J. Mitchell, W. R. Fright, B. C. McCallum, and T. R. Evans. Reconstruction and representation of 3d objects with radial basis functions. InACM SIGGRAPH, pages 67–76, 2001. 1

    work page 2001

  4. [4]

    Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe

    Rohan Chabra, Jan E. Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. InEuropean Conf. Computer Vision (ECCV), pages 608–625, 2020. 2

    work page 2020

  5. [5]

    Efficient geometry-aware 3d generative adversarial networks

    Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123–16133, 2022. 3

    2022

  6. [6]

    Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano- lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu

    Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano- lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model reposi- tory. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at ...

    arXiv 2015

  7. [7]

    Implicit feature net- works for texture completion from partial 3d data

    Julian Chibane and Gerard Pons-Moll. Implicit feature net- works for texture completion from partial 3d data. InEu- ropean Conference on Computer Vision (ECCV) Workshops, pages 717–725, 2020. 2

    work page 2020

  8. [8]

    Implicit functions in feature space for 3d shape reconstruc- tion and completion

    Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruc- tion and completion. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 6970–6981, 2020. 2, 6, 8

    work page 2020

Show all 54 references
  1. [9]

    Scan2mesh: From un- structured range scans to 3d meshes

    Angela Dai and Matthias Niessner. Scan2mesh: From un- structured range scans to 3d meshes. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 5569–5578, 2019. 2

    2019

  2. [10]

    Sg-nn: Sparse generative neural networks for self-supervised scene completion of rgb-d scans

    Angela Dai, Christian Diller, and Matthias Niessner. Sg-nn: Sparse generative neural networks for self-supervised scene completion of rgb-d scans. InIEEE/CVF Conf. Computer Vi- sion and Pattern Recognition (CVPR), pages 846–855, 2020. 2

    2020

  3. [11]

    Mitra, and Michael Wimmer

    Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J. Mitra, and Michael Wimmer. Points2surf: Learning implicit surfaces from point clouds. InEuropean Conf. Computer Vision (ECCV), pages 108–124, 2020. 2, 5, 6

    2020

  4. [12]

    Matthias Fey and Jan E. Lenssen. Fast graph representa- tion learning with PyTorch Geometric. InICLR Workshop on Representation Learning on Graphs and Manifolds, 2019. 6

    2019

  5. [13]

    Oriented-grid encoder for 3d implicit representations.Int’l Conf

    Arihant Gaur, G Dias Pais, and Pedro Miraldo. Oriented-grid encoder for 3d implicit representations.Int’l Conf. 3D Vision (3DV), 2024. 2

    2024

  6. [14]

    Kim, Bryan C

    Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mache ap- proach to learning 3d surface generation. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), pages 216–224, 2018. 2, 4

    2018

  7. [15]

    Blensor: Blender sensor simulation toolbox

    Michael Gschwandtner, Roland Kwitt, Andreas Uhl, and Wolfgang Pree. Blensor: Blender sensor simulation toolbox. InInternational Symposium on Visual Computing (ISVC), pages 199–208, 2011. 5

    2011

  8. [16]

    Probabilistic object detection: Definition and evaluation

    David Hall, Feras Dayoub, John Skinner, Haoyang Zhang, Dimity Miller, Peter Corke, Gustavo Carneiro, Anelia An- gelova, and Niko S¨underhauf. Probabilistic object detection: Definition and evaluation. InIEEE Winter Conf. on Applica- tions of Computer Vision (WACV), pages 1031–...

    2020

  9. [17]

    Point2mesh: A self-prior for deformable meshes.ACM Transactions on Graphics (TOG), 39(4), 2020

    Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen- Or. Point2mesh: A self-prior for deformable meshes.ACM Transactions on Graphics (TOG), 39(4), 2020. 8

    2020

  10. [18]

    Neural kernel surface recon- struction

    Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface recon- struction. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 4369–4379, 2023. 2, 6, 8

    2023

  11. [19]

    Batch normalization: Accelerating deep network training by reducing internal co- variate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal co- variate shift. InInt’l Conf. Machine learning (ICML), page 448–456, 2015. 5

    2015

  12. [20]

    Spatial transformer networks

    Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. InAd- vances in Neural Information Processing Systems (NIPS),

  13. [21]

    Local implicit grid representations for 3d scenes

    Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Niessner, and Thomas Funkhouser. Local implicit grid representations for 3d scenes. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 6000–6009, 2020. 2

    2020

  14. [22]

    Screened poisson sur- face reconstruction.ACM Transactions on Graphics (TOG), 32(3):1–13, 2013

    Michael Kazhdan and Hugues Hoppe. Screened poisson sur- face reconstruction.ACM Transactions on Graphics (TOG), 32(3):1–13, 2013. 1

    2013

  15. [23]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInt’l Conf. Learning Representa- tions (ICLR), 2015. 6

    2015

  16. [24]

    Abc: A big cad model dataset for geometric deep learning

    Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 9593...

    2019

  17. [25]

    Ravikrishna Kolluri, Jonathan Richard Shewchuk, and James F. O’Brien. Spectral surface reconstruction from noisy point clouds. InEurographics, pages 11–21, 2004. 1

    2004

  18. [26]

    Robust and efficient surface reconstruction from range data

    Patrick Labatut, Jean-Philippe Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. Computer Graphics Forum, 28(8):2275–2290, 2009. 1

    2009

  19. [27]

    Deep march- ing cubes: learning explicit surface representations

    Yiyi Liao, Simon Donn ´e, and Andreas Geiger. Deep march- ing cubes: learning explicit surface representations. In IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2916–2925, 2018. 2

    2018

  20. [28]

    Marching cubes: A high resolution 3d surface construction algorithm.ACM SIGGRAPH, 21(4):163–169, 1987

    William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm.ACM SIGGRAPH, 21(4):163–169, 1987. 3

    1987

  21. [29]

    Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representa- tion.ACM SIGGRAPH, 40(4), 2021. 2

    2021

  22. [30]

    Occupancy networks: Learning 3d reconstruction in function space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Se- bastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 4460–4470, 2019. 6

    2019

  23. [31]

    Meshry, D

    M. Meshry, D. B. Goldman, S. Khamis, H. Hoppe, R. Pandey, N. Snavely, and R. Martin-Brualla. Neural reren- dering in the wild. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 6871–6880, 2019

    2019

  24. [32]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. InEuropean Conf. Computer Vision (ECCV), pages 405–421, 2020. 2

    2020

  25. [33]

    Human-in-the-loop slam.Proceedings of the AAAI Conference on Artificial In- telligence (AAAI), 32(1), 2018

    Samer Nashed and Joydeep Biswas. Human-in-the-loop slam.Proceedings of the AAAI Conference on Artificial In- telligence (AAAI), 32(1), 2018. 1

    2018

  26. [34]

    Quadricslam: Dual quadrics from object detections as land- marks in object-oriented slam.IEEE Robotics and Automa- tion Letters (RA-L), 4(1):1–8, 2018

    Lachlan Nicholson, Michael Milford, and Niko S ¨underhauf. Quadricslam: Dual quadrics from object detections as land- marks in object-oriented slam.IEEE Robotics and Automa- tion Letters (RA-L), 4(1):1–8, 2018. 1

    2018

  27. [35]

    Deepsdf: Learning con- tinuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con- tinuous signed distance functions for shape representation. In IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 165–174, 2019. 2, 4

    2019

  28. [36]

    Py- torch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zem- ing Lin, Natalia Gimelshein, Luca Antiga, Alban Desmai- son, Andreas Kopf, Edward Yang, Zachary DeVito, Mar- tin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steine...

    2019

  29. [37]

    Convolutional occupancy networks

    Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. InEuropean Conf. Computer Vision (ECCV), pages 523–540, 2020. 2, 6, 8

    2020

  30. [38]

    Shape as points: A differentiable poisson solver

    Songyou Peng, Chiyu Max Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. InAdvances in Neu- ral Information Processing Systems (NeurIPS), 2021. 6, 8

    2021

  31. [39]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. InIEEE Conf. Computer Vision and Pat- tern Recognition (CVPR), pages 77–85, 2017. 2, 3, 5

    2017

  32. [40]

    Qi, Li Yi, Hao Su, and Leonidas J

    Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point- Net++: deep hierarchical feature learning on point sets in a metric space. InAdvances in Neural Information Processing Systems (NIPS), pages 5105–5114, 2017. 4

    2017

  33. [41]

    Pifu: Pixel-aligned implicit function for high-resolution clothed human digiti- zation

    Shunsuke Saito, , Zeng Huang, Ryota Natsume, Shigeo Mor- ishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digiti- zation. InIEEE/CVF Int’l Conf. Computer Vision (ICCV), pages 2304–2314, 2019. 2

    2019

  34. [42]

    Elhajj, and Daniel Asmar

    Abbas Sidaoui, Mohammad Kassem Zein, Imad H. Elhajj, and Daniel Asmar. A-slam: Human in-the-loop augmented slam. Inicra, pages 5245–5251, 2019. 1

    2019

  35. [43]

    Deep- voxels: Learning persistent 3d feature embeddings

    Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deep- voxels: Learning persistent 3d feature embeddings. In IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2437–2446, 2019. 2

    2019

  36. [44]

    Implicit neural representa- tions with periodic activation functions.Advances in Neural Information Processing Systems (NeurIPS), 33, 2020

    Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representa- tions with periodic activation functions.Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2

    2020

  37. [45]

    Laplacian surface editing

    Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rossl, and H-P Seidel. Laplacian surface editing. InEurographics, pages 175–184, 2004. 1

    2004

  38. [46]

    Neural geometric level of detail: Real-time rendering with implicit 3d shapes

    Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In IEEE/CVF Conf. Computer Vision and Pattern Recogni...

    2021

  39. [47]

    Fourier features let networks learn high frequency functions in low dimen- sional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ra- mamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimen- sional domains. InAdvances in Neural Informatio...

    2020

  40. [48]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems (NIPS), 2017. 2, 5

    2017

  41. [49]

    Pixel2mesh: Generating 3d mesh models from single rgb images

    Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. InEuropean Conf. Computer Vision (ECCV), pages 55–71, 2018. 2

    2018

  42. [50]

    Alto: Alternating latent topologies for implicit 3d reconstruction

    Zhen Wang, Shijie Zhou, Jeong Joon Park, Despoina Paschalidou, Suya You, Gordon Wetzstein, Leonidas Guibas, and Achuta Kadambi. Alto: Alternating latent topologies for implicit 3d reconstruction. InIEEE/CVF Conf. Computer Vi- sion and Pattern Recognition (CVPR), pages 259–270,...

    2023

  43. [51]

    Neural splines: Fitting 3d surfaces with infinitely- wide neural networks

    Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural splines: Fitting 3d surfaces with infinitely- wide neural networks. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 9949–9958, 2021. 2

    2021

  44. [52]

    Neural fields as learnable kernels for 3d reconstruction

    Francis Williams, Zan Gojcic, Sameh Khamis, Denis Zorin, Joan Bruna, Sanja Fidler, and Or Litany. Neural fields as learnable kernels for 3d reconstruction. InIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 18500–18510, 2022. 2

    2022

  45. [53]

    Mesh editing with poisson-based gradient field manipulation

    Yizhou Yu, Kun Zhou, Dong Xu, Xiaohan Shi, Hujun Bao, Baining Guo, and Heung-Yeung Shum. Mesh editing with poisson-based gradient field manipulation. InACM SIG- GRAPH, pages 644–651, 2004. 1

    2004

  46. [54]

    Thingi10k: A dataset of 10,000 3d-printing models.arXiv preprint arXiv:1605.04797, 2016

    Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models.arXiv preprint arXiv:1605.04797, 2016. 5, 8

    2016 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.