Pith. sign in

REVIEW 4 major objections 4 minor 50 references

High-Fidelity and Generalizable Neural Surface Reconstruction with Sparse Feature Volumes

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read SVRecon shows that a coarse occupancy field plus a sparse high-resolution feature volume pushes generalizable neural surface reconstruction to 512^3 resolution while cutting storage by more than 50-fold and improving accuracy.

desk verdict A benchmark-tested sparse-volume recipe that genuinely unlocks 512^3 GNSR on one GPU, with a real but partly under-stated dependency on first-stage occupancy recall. read the letter →

arxiv 2507.05952 v1 pith:ONRQTL6Y submitted 2025-07-08 cs.CV

classification cs.CV
keywords SVReconsparsefeaturevolumesgeneralizableneuralsurfacereconstructionoccupancypredictionvolumerenderingtransformermulti-viewconvolution
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

SVRecon is a two-stage pipeline for generalizable neural surface reconstruction that replaces the dense 3D feature volume used by earlier methods with a sparse one. A first network predicts occupancy at a coarse $128^3$ resolution from posed images and depth maps; a second network builds a $512^3$ feature volume only inside the occupied voxels and renders surfaces with a transformer-based volume renderer. The authors report that this lowers storage by more than $50\times$, retains only about 1.9% of voxels, and improves reconstruction accuracy on the DTU benchmark, matching the best Chamfer distance while exceeding all compared methods in normal consistency. The approach also transfers to BlendedMVS and Tanks and Temples without retraining. If correct, it shows that the resolution ceiling in generalizable neural surface reconstruction is a representation problem rather than a learning problem.

What carries the argument

The central object is the nested sparse feature volume: a coarse occupancy grid at $128^3$ predicted by a 3D U-Net from mean-variance multi-view features, supersampled by a factor of $s=4$ into sparse high-resolution volumes that store features only in occupied coarse voxels. A dense lookup table maps any query point's occupancy coordinates to its mini-volume index, giving $O(1)$ trilinear interpolation without densifying the volume. The sparse UNet aggregates features globally, and a transformer volume renderer (Eq. 5 in the paper) turns per-ray feature aggregations into color and depth. This mechanism is what decouples memory from scene volume: storage scales with surface area, not with the bounding box at full resolution.

What would settle it

Take a scene whose evaluation region contains a large textureless planar surface, such as a table or wall, and measure occupancy recall at $128^3$; if the coarse stage with threshold $\tau=0.1$ and $3\times3\times3$ dilation misses that surface, the sparse $512^3$ stage has no occupied voxels there, so the final mesh will contain a hole regardless of the high-resolution stage. This directly tests whether the 96.8% recall claim holds outside the DTU protocol.

Watch

Extended reading notes

Core claim

The central discovery is that a nested two-layer volume—a coarse binary occupancy field at $128^3$ plus sparse high-resolution features at $512^3$ inside occupied voxels—carries enough information for high-fidelity, generalizable surface reconstruction from a few images. The occupancy stage is trained with focal loss on ground-truth surfaces and binarized with a low threshold plus dilation to maximize recall (96.8% on DTU), so no surface voxel of importance is discarded. The sparse stage reimplements ray sampling, feature aggregation, and trilinear querying over sparse volumes using a low-resolution dense lookup table that keeps query cost $O(1)$. On the DTU test set the method achieves a mean Chamfer distance of 1.00, matching the best baseline, and a normal-consistency AUC up to 15 degrees of 21.00, ahead of all compared methods, at $512^3$ resolution on 32GB GPUs.

Load-bearing premise

The entire pipeline stands on the coarse $128^3$ occupancy stage recalling every surface voxel that matters, because any voxel it labels empty is permanently absent from the sparse high-resolution volume and cannot be recovered later.

Editorial extensions

If this is right

  • Generalizable reconstruction on standard GPUs can move from $128^3$ to $512^3$ effective resolution, because the sparse volume keeps memory proportional to surface area rather than scene volume.
  • Reconstruction quality rises with effective resolution: the paper's ablation shows mean Chamfer distance improving from 1.27 at $128^3$ to 1.04 at $256^3$ to 1.00 at $512^3$.
  • The same pretrained model transfers to new datasets (BlendedMVS, Tanks and Temples) without per-scene optimization, so the two-stage sparsity does not sacrifice generalization.
  • Using more input views improves results, and the mean-variance feature construction makes the network indifferent to the exact view count.
  • Because only occupied voxels are stored, storage savings grow with scene sparsity; the measured 1.89% space efficiency implies more than a $50\times$ reduction for typical surfaces.

Reading between the lines

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

  • We infer that the same coarse-occupancy, supersampling scheme could be extended to unbounded scenes by replacing the fixed $128^3$ bounding-box grid with an adaptive coarse structure, since the heavy computation happens only in occupied regions.
  • We infer that the dense lookup table used for $O(1)$ sparse queries has memory $O(K^3)$ even though the feature volume is sparse, so at resolutions beyond $512^3$ the lookup table, not the features, would become the memory bottleneck; a hierarchical lookup would be the natural fix.
  • We infer that the normal-consistency gain over a baseline with equal Chamfer distance indicates the resolution increase mostly improves surface orientation and smoothness rather than point proximity; a testable prediction is that normal-error drops faster than Chamfer distance as the supersampling factor grows.
  • We infer that a stress-test scene with a large untextured plane inside the reconstruction volume would expose the coarse occupancy stage's recall limit, because a missed voxel at $128^3$ cannot be recovered by the fine stage.
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 / 4 minor

Summary. The paper proposes SVRecon, a two-stage generalizable neural surface reconstruction method. Stage one predicts a coarse 128^3 occupancy field from posed images via a 3D U-Net; stage two supersamples each occupied coarse voxel by s=4, builds sparse 512^3 feature volumes only inside the occupied mask, and performs volume rendering with a transformer using custom sparse sampling and querying. The authors report that the sparse representation occupies about 1.89% of voxels (about 50x storage saving), enables 512^3 resolution on a 32GB GPU where dense methods are limited to 128^3, and achieves Chamfer distance comparable to or better than prior generalizable methods while substantially improving normal consistency on DTU. Generalization to BlendedMVS and Tanks and Temples is demonstrated qualitatively, and ablation studies examine resolution, view count, and feature channels.

Significance. If the claims hold, the paper makes a useful practical contribution: it shows that a simple two-layer sparse volume, rather than an octree or dense grid, can lift generalizable neural surface reconstruction to 512^3 resolution on standard hardware while retaining cross-scene generalization. The reported 50x storage reduction and the large improvement in normal consistency over ReTR and UFORecon are concrete, measurable gains on a public benchmark. The custom sparse lookup and ray-sampling algorithms are a plausible engineering contribution. The main reservation is that the entire method is gated by the coarse occupancy mask, and the reported 96.8% recall is measured on a DTU protocol that excludes the textureless table; the generalizability and accuracy claims therefore need to be qualified and tested more adversarially. The paper ships quantitative comparisons using released baseline code, which strengthens reproducibility, but no code release for the proposed method is mentioned.

major comments (4)
  1. [Sec. 3.2-3.4 and Sec. 4.2] The two-stage pipeline irreversibly gates the final reconstruction on the coarse occupancy mask eO = dilate(I(O >= tau)): any true surface voxel outside eO is never represented in the sparse 512^3 volume and cannot be recovered by the second stage. Table 4 reports mean recall 96.8%, but per-scan recall drops to 90.4% (scan 40), and the authors state that missed geometry is mostly the textureless table, which is excluded from the standard DTU evaluation. Because tau = 0.1 and the 3x3x3 dilation are hand-tuned on DTU to maximize recall, the 96.8% figure is not established as a transferable property; scenes with textureless walls or floors inside the evaluation region, or surfaces thinner than a 128^3 coarse voxel, could lose geometry irrecoverably. The paper should report sensitivity of final reconstruction quality to tau and dilation, and evaluate recall and final metrics on protocols that do not exclude the table region or on additional scenes with textureless surfaces.
  2. [Abstract and Sec. 4.1, Tables 2-3] The abstract claims 'superior reconstruction accuracy over current state-of-the-art methods,' but Table 2 shows mean Chamfer distance 1.00 for both SVRecon and UFORecon, so on the standard metric the methods are tied. The accuracy advantage rests entirely on Normal Consistency in Table 3. The claim should be qualified, for example as 'comparable Chamfer distance with substantially better normal consistency,' or the authors should justify why NC should be treated as the primary accuracy measure in this setting.
  3. [Sec. 4.3, Fig. 5] The generalization claim is supported only by qualitative visualizations on BlendedMVS and Tanks and Temples, with no quantitative metrics or baseline comparisons on those datasets. Given that the title claims generalizability and that the occupancy-gating failure mode is scene-content dependent, the paper needs quantitative cross-dataset evaluation (CD/NC, or at least occupancy recall against ground truth) to substantiate that the method transfers beyond DTU.
  4. [Sec. 4.2, Table 4 and Supplementary Sec. 6] The 'more than 50x storage reduction' is derived from the 1.89% space efficiency of the occupied high-resolution voxels, but the dense K^3 lookup table H used for sparse querying and the dense 128^3 first-stage feature volume are not included in this ratio. The storage claim should specify exactly which tensors are counted, and whether the overhead of H (K=128) and the first-stage volume is amortized or neglected.
minor comments (4)
  1. [Throughout] There are numerous typos and grammatical slips, including 'trianing' (Sec. 3.2), 'V olumetric' in the title header, 'compated' (Conclusion), 'intuitition' (Sec. 4.3), 'syntheiss' (Sec. 4.3), 'thresold' (Supplementary Sec. 7), and 'as is a standard metric' (Sec. 4.1). A careful proofread is needed.
  2. [Sec. 4.1 and Implementation Details] The abstract says 'consumer-grade GPUs with 32GB of VRAM,' but training uses 4 A100 GPUs and about 30GB memory; please clarify that the consumer-grade claim applies to testing/inference and specify the inference hardware.
  3. [Fig. 3 and Fig. 5] The red rectangles marking detail regions are small and may not be visible in print; consider zoomed insets to make the qualitative differences easier to verify.
  4. [Sec. 4.3, Novel View Synthesis] The statement that the background is not modeled is an important limitation and should appear in the main text limitations discussion, not only as a passing remark in the novel-view synthesis paragraph.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the sparse-volume reconstruction is evaluated on held-out DTU scenes against external baselines, and the coarse-occupancy threshold and dilation are disclosed hyperparameter choices rather than self-derived predictions.

full rationale

The paper's derivation chain is: predict a coarse occupancy field O at 128^3 from multi-view features; binarize and dilate it to eO = dilate(I(O >= tau)); supersample occupied voxels by s=4 to build a sparse 512^3 feature volume; perform volume rendering only inside occupied fragments; and evaluate the resulting surfaces on held-out DTU test scenes using Chamfer distance and normal consistency against external baselines such as ReTR, UFORecon, and VolRecon. No load-bearing step reduces to its own input. The occupancy threshold tau=0.1 and the 3x3x3 dilation are hand-set to maximize recall and are fully disclosed in Section 4.2 and Supplementary Section 7; tuning hyperparameters on a benchmark is standard practice, not a fitted parameter being renamed as a prediction. The reported 96.8% recall and 1.89% space efficiency are empirical measurements on held-out scenes, not quantities that are forced by construction. The >50x storage reduction follows arithmetically from the measured occupancy fraction (1/0.0189), so it is not circular. The rendering formulation is explicitly adopted from ReTR [21], an external baseline, and the authors' own self-citation (Fua 1997, reference [6]) is a general multi-view-stereo reference that is not load-bearing. The acknowledged limitation that coarse occupancy can miss textureless geometry excluded from the standard DTU protocol is a genuine generalization risk, but it is a limitation of the method, not circular reasoning. The central claim is therefore self-contained with respect to its evaluation protocol and external baselines.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard supervised learning assumptions plus one domain-specific bottleneck: the coarse occupancy field must recall the surface. The free parameters are hyperparameters of the pipeline, chiefly tau, the dilation, the supersampling factor, and the unreported depth loss weight alpha. No new physical entities are introduced; the dense lookup table is an algorithmic data structure, not a postulated entity.

free parameters (5)
  • occupancy threshold tau = 0.1
    Hand-chosen conservative threshold to binarize occupancy predictions; controls the trade-off between precision and recall and therefore which voxels are kept for high-resolution reconstruction (Sec. 3.2 and Sec. 4.2).
  • dilation kernel size and rethreshold = 3x3x3 kernel, rethreshold at 33*0.1
    Morphological dilation applied to predicted occupancy to maximize recall; hand-tuned and directly inflates space efficiency from 0.45% to 1.89% (Sec. 4.2 and supplementary Sec. 7).
  • supersampling factor s = 4
    Sets the effective high-resolution grid to (sK)^3 = 512^3; chosen because memory prevents higher values (Sec. 3.3 and Sec. 4.3).
  • depth loss weight alpha = not reported
    Balance term between color and depth losses in Eq. (6); the paper does not state its value, leaving a gap in reproducibility.
  • occupancy resolution K = 128
    Voxel resolution of the coarse stage; memory-constrained choice (Sec. 3.2).
assumptions (4)
  • domain assumption Per-channel mean and variance of multi-view image features are a sufficient 3D feature statistic
    Eq. (2) reduces M view features to MeanVar before 3D aggregation; this information bottleneck is assumed to preserve geometry-relevant content.
  • domain assumption Ground-truth occupancies can be derived by merging per-view ground-truth depth point clouds
    Used to supervise the occupancy network (Sec. 3.2); assumes merged depth points cover all visible surfaces and that occluded ground-truth points should be excluded.
  • domain assumption The coarse 128^3 occupancy field with dilation can recall all geometry relevant for the benchmark
    The two-stage pipeline cannot recover voxels missed at stage one; the paper relies on 96.8% recall on DTU, with missed geometry confined to regions excluded from standard evaluation (Sec. 4.2).
  • standard math Surface regions occupy a small fraction of voxels, so sparsity does not hurt reconstruction given perfect occupancy
    Surfaces have measure zero, so a sufficient occupancy oracle would keep only relevant voxels; the practical validity depends on the occupancy network (Sec. 1 and Sec. 4.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Fidelity and Generalizable Neural Surface Reconstruction with Sparse Feature Volumes." pith.science (2026). https://pith.science/paper/ONRQTL6Y

@misc{pith2026250705952,
  author       = {Pith},
  title        = {Pith review of: High-Fidelity and Generalizable Neural Surface Reconstruction with Sparse Feature Volumes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ONRQTL6Y}},
  note         = {Machine review of arXiv:2507.05952}
}
abstract

Generalizable neural surface reconstruction has become a compelling technique to reconstruct from few images without per-scene optimization, where dense 3D feature volume has proven effective as a global representation of scenes. However, the dense representation does not scale well to increasing voxel resolutions, severely limiting the reconstruction quality. We thus present a sparse representation method, that maximizes memory efficiency and enables significantly higher resolution reconstructions on standard hardware. We implement this through a two-stage approach: First training a network to predict voxel occupancies from posed images and associated depth maps, then computing features and performing volume rendering only in voxels with sufficiently high occupancy estimates. To support this sparse representation, we developed custom algorithms for efficient sampling, feature aggregation, and querying from sparse volumes-overcoming the dense-volume assumptions inherent in existing works. Experiments on public datasets demonstrate that our approach reduces storage requirements by more than 50 times without performance degradation, enabling reconstructions at $512^3$ resolution compared to the typical $128^3$ on similar hardware, and achieving superior reconstruction accuracy over current state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2507.05952 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. SVRecon pipeline. First, we build a low-resolution dense 3D feature volume from multi-view features and predict 3D occupan￾cies. Second, we construct high-resolution sparse feature volumes where the predicted occupancy warrants it. Finally, volume rendering Transformers are used to aggregate per-ray sampled features and infer color and depth, enabling scene reconstruction. input images, a number of generalizable app… view at source ↗
Figure 3
Figure 3. Sparse-view surface reconstructions on DTU test scenes. In the middle columns, we show shaded surfaces for UFORecon, ReTR, and our approach. In the rightmost column, we show the occupancies predicted by the first stage of our method at resolution 1283 . Voxels are shrunk slightly for visualization purposes. The red rectangles highlight the region with visible differences. Scan Mean 24 37 40 55 63 65 69 83 97 105 106… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Normal Consistency evaluation on scan24 from the DTU dataset. The normal differences are visualized using colors. Errors in the range [0 ◦ , 15◦ ] are color coded linearly from white to red. Error larger than 15◦ are shown in green. Our SVRecon method significantly out…
Figure 6
Figure 6. Figure 6: Novel view synthesis examples of our SVRecon method on scan24 and scan105 from the DTU dataset. The details are pre￾served well for the foreground object. 5. Conclusion and Future Work In this paper, we propose a two-stage neural surface recon￾struction method based on…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 44 canonical work pages

  1. [1]

    Large-scale data for multiple-view stereopsis

    Henrik Aanæs, Rasmus Ramsbøl Jensen, George V ogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 120:153–168, 2016. 5

  2. [2]

    Mvsformer++: Revealing the devil in transformer’s details for multi-view stereo

    Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer++: Revealing the devil in transformer’s details for multi-view stereo. arXiv preprint arXiv:2401.11673, 2024. 2

  3. [3]

    Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124–14133, 2021. 6

  4. [4]

    A volumetric method for building complex models from range images

    Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Conference on Computer graphics and interactive techniques , pages 303–312, 1996. 5

  5. [5]

    Transmvsnet: Global context-aware multi-view stereo network with trans- formers

    Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, and Xiao Liu. Transmvsnet: Global context-aware multi-view stereo network with trans- formers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8585–8594,

  6. [6]

    P. Fua. From Multiple Stereo Views to Multiple 3D Sur- faces. International Journal of Computer Vision, 24(1):19– 35, 1997. 2

  7. [7]

    Accurate, dense, and robust multiview stereopsis

    Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence , 32(8):1362–1376, 2009. 2

  8. [8]

    Cascade cost volume for high-resolution multi-view stereo and stereo matching

    Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495–2504, 2020. 2

Show all 50 references
  1. [9]

    Hartley and A

    R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 2

  2. [10]

    2d gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1–11, 2024. 1, 3

  3. [11]

    Surfacenet: An end-to-end 3d neural network for mul- tiview stereopsis

    Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for mul- tiview stereopsis. In ICCV, pages 2307–2315, 2017. 2

  4. [12]

    Sur- facenet+: An end-to-end 3d neural network for very sparse multi-view stereopsis

    Mengqi Ji, Jinzhi Zhang, Qionghai Dai, and Lu Fang. Sur- facenet+: An end-to-end 3d neural network for very sparse multi-view stereopsis. IEEE Transactions on Pattern Analy- sis and Machine Intelligence, 43(11):4078–4093, 2020

  5. [13]

    A. Kar, C. H ¨ane, and J. Malik. Learning a Multi-View Stereo Machine. In Advances in Neural Information Processing Systems, pages 364–375, 2017. 2

  6. [14]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), 2023. 1, 3

  7. [15]

    D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In arXiv Preprint, 2014. 5

  8. [16]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM TOG, 36(4), 2017. 5, 7

  9. [17]

    Kostrikov and J

    I. Kostrikov and J. Gall. Depth Sweep Regression Forests for Estimating 3D Human Pose from Images. InBritish Machine Vision Conference, 2014. 2

  10. [18]

    Kutulakos and S.M

    K.N. Kutulakos and S.M. Seitz. A Theory of Shape by Space Carving. International Journal of Computer Vision , 38(3): 197–216, 2000

  11. [19]

    A quasi-dense approach to surface reconstruction from uncalibrated images

    Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence , 27(3):418–433, 2005. 2

  12. [20]

    Z. Li, T. M ¨uller, A. Evans, R. Taylor, M. Unberath, M. Liu, and C. Lin. Neuralangelo: High-Fidelity Neural Surface Re- construction. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2

  13. [21]

    Retr: Modeling rendering via transformer for generalizable neural surface re- construction

    Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface re- construction. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 5, 6

  14. [22]

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie. Feature Pyramid Networks for Object Detection. In Conference on Computer Vision and Pattern Recognition,

  15. [23]

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll´ar. Focal Loss for Dense Object Detection. In International Confer- ence on Computer Vision, 2017. 4

  16. [24]

    Neural sparse voxel fields

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651–15663,

  17. [25]

    Sparseneus: Fast generalizable neural sur- face reconstruction from sparse views

    Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural sur- face reconstruction from sparse views. In European Confer- ence on Computer Vision, pages 210–227. Springer, 2022. 1, 2, 3, 4, 6

  18. [26]

    Lorensen and H.E

    W.E. Lorensen and H.E. Cline. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. In ACM SIGGRAPH, pages 163–169, 1987. 5

  19. [27]

    Ben Mildenhall, S. P. P., M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng. NeRF: Representing Scenes as Neu- ral Radiance Fields for View Synthesis. In European Con- ference on Computer Vision, 2020. 1, 2

  20. [28]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1–15, 2022. 2

  21. [29]

    Uforecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets

    Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. Uforecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094–5104, 2024. 1,...

  22. [30]

    V olrecon: V olume rendering of signed ray distance functions for generalizable multi-view recon- struction

    Yufan Ren, Fangjinhua Wang, Tong Zhang, Marc Pollefeys, and Sabine S¨usstrunk. V olrecon: V olume rendering of signed ray distance functions for generalizable multi-view recon- struction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pa...

  23. [31]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-Net: Convo- lutional Networks for Biomedical Image Segmentation. In Conference on Medical Image Computing and Computer As- sisted Intervention, pages 234–241, 2015. 4

  24. [32]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes L Sch ¨onberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, pages 501–518,

  25. [33]

    Seitz, B

    S.M. Seitz, B. Curless, J. Diebel, D. Scharstein, and R. Szeliski. A Comparison and Evaluation of Multi-View Stereo Reconstruction Algorithms. In Conference on Com- puter Vision and Pattern Recognition, pages 519–528, 2006. 2

  26. [34]

    Shum and S

    H. Shum and S. B. Kang. Review of Image-Based Rendering Techniques. In Visual Communications and Image Process- ing, pages 2–13, 2000. 2

  27. [35]

    Torchsparse++: Efficient training and inference framework for sparse convolution on gpus

    Haotian Tang, Shang Yang, Zhijian Liu, Ke Hong, Zhong- ming Yu, Xiuyu Li, Guohao Dai, Yu Wang, and Song Han. Torchsparse++: Efficient training and inference framework for sparse convolution on gpus. In Proceedings of the 56th Annual IEEE/ACM International Symposium on Microarc...

  28. [36]

    Itermvs: Iterative probability estimation for efficient multi-view stereo

    Fangjinhua Wang, Silvano Galliani, Christoph V ogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8606–8615, 2022. 2

  29. [37]

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wang. Neus: Learning Neural Implicit Surfaces by V olume Rendering for Multi-View Reconstruction. In Advances in Neural Information Processing Systems, 2021. 1, 2, 6

  30. [38]

    Ibr- net: Learning multi-view image-based rendering

    Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibr- net: Learning multi-view image-based rendering. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and ...

  31. [39]

    C2f2neus: Cascade cost frustum fusion for high fidelity and generalizable neu- ral surface reconstruction

    Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhao- jie Zeng, Junle Wang, and Wei Yang. C2f2neus: Cascade cost frustum fusion for high fidelity and generalizable neu- ral surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, page...

  32. [40]

    Mvsnet: Depth inference for unstructured multi-view stereo

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vi- sion (ECCV), pages 767–783, 2018. 2

  33. [41]

    Recurrent mvsnet for high-resolution multi-view stereo depth inference

    Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5525–5534, 2019. 2

  34. [42]

    Blendedmvs: A large- scale dataset for generalized multi-view stereo networks

    Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large- scale dataset for generalized multi-view stereo networks. In CVPR, pages 1790–1799, 2020. 5, 7

  35. [43]

    Yariv, Y

    L. Yariv, Y . Kasten, D. Moran, M. Galun, M. Atzmon, B. Ronen, and Y . Lipman. Multiview Neural Surface Recon- struction by Disentangling Geometry and Appearance. In Advances in Neural Information Processing Systems , 2020. 1, 2

  36. [44]

    V olume rendering of neural implicit surfaces

    Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. V olume rendering of neural implicit surfaces. In Thirty- Fifth Conference on Neural Information Processing Systems,

  37. [45]

    Spar- secraft: Few-shot neural reconstruction through stereopsis guided geometric linearization

    Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Spar- secraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37–56. Springer, 2024. 1, 2, 6

  38. [46]

    pixelnerf: Neural radiance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578–4587, 2021. 6

  39. [47]

    Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM TOG, 43(6):1–13, 2024. 1, 3

  40. [48]

    Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers

    Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. arXiv Preprint, 2024. 3 10 High-Fidelity and Generalizable Neural Surface Reconstruction with Sparse Featur...

  41. [49]

    Querying Sparse V olumes. Hereafter, we will use coarse-voxel to indicate a voxel at occupancy prediction resolution K 3, and fine-voxel to in- dicate a mini-voxel in each mini-volume effectively at res- olution (sK)3. To explain the algorithm, we first define a global grid fr...

  42. [50]

    Network Architectures

    More Implementation Details. Network Architectures. We use Feature Pyramid Net- work (FPN) for image feature extraction. To obtain pro- jection features from 3D points, we interpolate on the 1/2 resolution feature map from FPN. For 3D sparse fea- ture volumes we use a sparse U...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.