arxiv: 2604.04925 · v2 · submitted 2026-04-06 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

SimpleProc: Fully Procedural Synthetic Data from Simple Rules for Multi-View Stereo

Zeyu Ma , Alexander Raistrick , Jia Deng

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:58 UTC · model grok-4.3

classification 💻 cs.CV

keywords synthetic dataprocedural generationmulti-view stereo3D reconstructioncomputer visiontraining datasetsNURBS surfaces

0 comments

The pith

Procedural generation with a few simple rules produces training data for multi-view stereo that rivals much larger sets of manually curated images.

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

The paper introduces SimpleProc, a fully procedural method to generate synthetic training data for multi-view stereo using only a small set of rules. These rules rely on NURBS surfaces combined with basic displacement and texture patterns to create varied 3D scenes. At a modest scale of 8,000 images, models trained on this synthetic data already outperform those trained on the same number of manually curated images from games and real objects. Scaling the procedural dataset to 352,000 images yields models whose performance matches or exceeds that of models trained on over 692,000 curated images across multiple benchmarks.

Core claim

SimpleProc is a fully procedural generator for multi-view stereo training data driven by a small set of rules based on Non-Uniform Rational Basis Splines (NURBS), displacement maps, and texture patterns. This approach generates entirely synthetic scenes without any manual curation or real-world capture. Experiments show that training on 8,000 SimpleProc images yields better results than training on 8,000 manually curated images. At larger scale, 352,000 procedural images enable models to achieve comparable or superior accuracy on MVS benchmarks compared to training on 692,000 curated images.

What carries the argument

The SimpleProc generator, which constructs 3D scenes from NURBS surfaces modified by simple displacement and texture rules before rendering multi-view image sets.

If this is right

Larger synthetic datasets can be produced efficiently to improve MVS model performance without the cost of manual data collection.
The performance advantage holds when scaling up the number of training images.
Synthetic data from simple rules can replace or supplement curated real and game-derived datasets for stereo reconstruction tasks.
The method demonstrates that limited rule sets suffice to capture useful scene variety for this task.

Where Pith is reading between the lines

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

Similar procedural approaches might reduce data requirements for other vision tasks that rely on 3D structure.
Further refinement of the rules could address any remaining performance gaps on specific benchmarks.
This work opens the possibility of generating unlimited training data tailored to particular environments or object types.

Load-bearing premise

The diversity and realism of scenes produced by the limited procedural rules must match the statistical properties of real multi-view stereo data closely enough to transfer learning gains.

What would settle it

Training a model on the scaled SimpleProc dataset and testing it on a broad set of real-world MVS benchmarks where it underperforms models trained on equivalent curated data would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.04925 by Alexander Raistrick, Jia Deng, Zeyu Ma.

**Figure 1.** Figure 1: Fully procedural synthetic data from simple rules (top) is as effective as curated data from artists or 3D scans (bottom) for training multi-view stereo models. arXiv:2604.04925v2 [cs.CV] 7 Apr 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Figure 2: Procedural Data Generation Pipeline. Stage 1 generates NURBS surfaces from the lofting operation. Stage 2 applies displacements, textures, and material properties. Stage 3 arranges cameras, objects, lighting, and optional room boxes. 4 Data Generation Pipeline Overview. Our data generation pipeline ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Shape generation pipeline. Top: Profile curves (Starfish and Reptile styles) with control points in red (only one of the profiles is shown for each shape). Middle: Stem curves from 3D random walks. Bottom: Resulting lofted NURBS surfaces showing diverse smooth and sharp features [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Shapes with Displacement and Textures. The top row shows displacements applied to base shapes. The middle row shows textures from brick, wave, or Perlin noise pattern. The bottom row shows complex textures created by boolean operations. 4.4 Camera Placement To make the objects more efficiently used, we place cameras before adding objects into the scene. The position of the camera is sampled by randomly sel… view at source ↗

**Figure 5.** Figure 5: Gallery of 10 random examples of our generated scenes the MegaSynth [12] large-scale synthetic dataset. We maximize the number of unique scenes to ensure the greatest diversity. A minimum of 8 images per scene is required for MVSAnywhere training, resulting in 1000 unique scenes. When sampling from the eight-dataset mixture, we maintain the original distribution ratio of the constituent datasets. Since Meg… view at source ↗

**Figure 6.** Figure 6: Unlimited-Budget qualitative comparison across five benchmarks. Our procedural training data yields competitive depth estimation results compared to the 8-dataset baseline. model trained on 8-dataset and the model trained on our data. In each example we have competitive results and it is easy to find a region where we perform better (in red boxes). Note that there are circle/stripe patterns in the error m… view at source ↗

read the original abstract

In this paper, we explore the design space of procedural rules for multi-view stereo (MVS). We demonstrate that we can generate effective training data using SimpleProc: a new, fully procedural generator driven by a very small set of rules using Non-Uniform Rational Basis Splines (NURBS), as well as basic displacement and texture patterns. At a modest scale of 8,000 images, our approach achieves superior results compared to manually curated images (at the same scale) sourced from games and real-world objects. When scaled to 352,000 images, our method yields performance comparable to--and in several benchmarks, exceeding--models trained on over 692,000 manually curated images. The source code and the data are available at https://github.com/princeton-vl/SimpleProc.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Simple NURBS rules generate competitive MVS data at scale but the generalization depends on unverified distribution similarity.

read the letter

The one thing to know is that this work shows a minimal procedural generator using NURBS can produce synthetic MVS training data that scales to compete with much larger hand-curated collections. They introduce SimpleProc with a small set of rules for surfaces, displacements, and textures. At 8k images it already does better than manual data at the same scale from games and real objects. Scaling to 352k images gets results on par with or better than 692k manual images on several benchmarks. The code and data release is helpful for anyone wanting to try it. The approach is new in its extreme simplicity for this task. Most synthetic MVS pipelines use more complex modeling or mix in real data. Here the focus on fully procedural keeps things cheap and controllable. The soft spot is in the lack of evidence that the generated data matches real MVS distributions. Simple rules are likely to miss some natural variations in geometry and illumination that matter for stereo matching. No stats on how the synthetic scenes compare to real ones in terms of depth ranges or view overlaps are mentioned, so the performance edge could be specific to the benchmarks rather than a general win. The abstract is light on experimental details like exact splits and metrics, which makes the claims harder to assess fully. This is useful for vision researchers looking at data-efficient training for 3D tasks. It deserves peer review because the results are concrete, the method is reproducible, and it raises a practical question about how simple synthetic data can be while still being effective. A referee could push for more distribution analysis, but the paper is solid enough to review.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces SimpleProc, a fully procedural synthetic data generator for multi-view stereo (MVS) driven by a minimal set of rules based on NURBS surfaces, basic displacement maps, and simple texture patterns. It reports that models trained on 8,000 procedurally generated images outperform those trained on the same number of manually curated images from games and real objects, and that scaling the procedural data to 352,000 images produces performance comparable to—and in several cases exceeding—models trained on over 692,000 manually curated images across multiple benchmarks. Code and data are released publicly.

Significance. If the empirical scaling results hold under closer scrutiny, the work provides evidence that simple, fully procedural rules can generate training data whose quality rivals large-scale manually curated collections for MVS, potentially lowering the barrier to high-performance models. The release of code and data strengthens reproducibility.

major comments (2)

[Experiments section] Experiments section: the headline scaling claim (352k procedural images matching/exceeding 692k manual) is load-bearing, yet the manuscript provides no quantitative distributional diagnostics (e.g., histograms of surface normals, depth-gradient statistics, or view-overlap entropy) comparing the procedural output to real MVS datasets such as DTU or Tanks & Temples. Without this, it remains unclear whether gains are attributable to scale or to unintended alignment between the limited NURBS+displacement rules and the evaluation distributions.
[Experiments section] Experimental protocol: the abstract and results report clear scale comparisons and performance numbers, but details on exact benchmark splits, metric definitions, training hyperparameters, and controls for distribution shift between procedural and manual data are insufficiently specified. This weakens the support for the central claim that procedural data is generally sufficient.

minor comments (2)

[Abstract] Abstract: the phrase 'in several benchmarks' should name the specific datasets (e.g., DTU, Tanks & Temples) to allow immediate assessment of scope.
[Method] Method description: the precise parameterization of the NURBS surfaces and displacement functions could be expanded with a short table of default ranges to aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comments point by point below, proposing specific revisions to the Experiments section to strengthen the presentation of our results.

read point-by-point responses

Referee: [Experiments section] Experiments section: the headline scaling claim (352k procedural images matching/exceeding 692k manual) is load-bearing, yet the manuscript provides no quantitative distributional diagnostics (e.g., histograms of surface normals, depth-gradient statistics, or view-overlap entropy) comparing the procedural output to real MVS datasets such as DTU or Tanks & Temples. Without this, it remains unclear whether gains are attributable to scale or to unintended alignment between the limited NURBS+displacement rules and the evaluation distributions.

Authors: We agree that quantitative distributional diagnostics would help clarify the source of the observed performance gains. In the revised manuscript, we will add a new subsection in Experiments that includes histograms and statistics for surface normals, depth gradients, and view-overlap entropy, computed on samples from SimpleProc and directly compared against the DTU and Tanks & Temples datasets. These diagnostics will be generated from the publicly released code and data to demonstrate that the procedural distribution is broad and does not exhibit unintended alignment with the evaluation sets. revision: yes
Referee: [Experiments section] Experimental protocol: the abstract and results report clear scale comparisons and performance numbers, but details on exact benchmark splits, metric definitions, training hyperparameters, and controls for distribution shift between procedural and manual data are insufficiently specified. This weakens the support for the central claim that procedural data is generally sufficient.

Authors: We acknowledge that the current manuscript lacks sufficient detail on the experimental protocol. We will expand the Experiments section with: (1) exact benchmark splits, including which specific scenes or subsets from DTU, Tanks & Temples, and other evaluation sets are used for training, validation, and testing; (2) precise definitions of all reported metrics (e.g., accuracy, completeness, and any custom thresholds); (3) complete training hyperparameters, such as optimizer settings, learning rate schedules, batch sizes, number of epochs, and data augmentation procedures; and (4) explicit controls for distribution shift, including verification that procedural scenes do not overlap with evaluation scenes in geometry or object categories. These additions will make the protocol fully reproducible and provide stronger support for the generality of procedural data. revision: yes

Circularity Check

0 steps flagged

No circularity detected in empirical claims or generator design

full rationale

The paper presents an empirical demonstration: explicit procedural rules (NURBS surfaces plus basic displacement and texture patterns) are used to synthesize training images, models are trained on the resulting data, and performance is measured on external benchmarks (DTU, Tanks & Temples, etc.) against models trained on independent manually-curated datasets. No derivation, equation, or scaling claim reduces by construction to the input rules; the performance numbers are direct experimental outcomes rather than fitted parameters renamed as predictions. No self-citation load-bearing steps, uniqueness theorems, or ansatzes are invoked to justify the central result. The work is self-contained as a standard data-generation experiment.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that a small set of NURBS and pattern rules can produce training data whose distribution supports strong generalization to real MVS tasks; no explicit free parameters or invented entities are described in the abstract.

axioms (1)

domain assumption A small set of procedural rules using NURBS, displacement, and textures generates scenes with sufficient variety and realism for effective MVS training
Invoked to justify why the generated data outperforms or matches manual curation.

pith-pipeline@v0.9.0 · 5435 in / 1203 out tokens · 45149 ms · 2026-05-10T18:58:27.356565+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

shapes are generated by lofting a profile curve through a stem curve... NURBS surfaces... Perlin noise... brick, wave, or noise texture
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

When scaled to 352,000 images, our method yields performance comparable to... over 692,000 manually curated images

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

31 extracted references · 5 canonical work pages · 2 internal anchors

[1]

International Journal of Computer Vision (IJCV)120(2), 153–168 (2016)

Aanæs, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. International Journal of Computer Vision (IJCV)120(2), 153–168 (2016)

2016
[2]

Virtual KITTI 2

Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2. arXiv preprint arXiv:2001.10773 (2020)

work page internal anchor Pith review arXiv 2001
[3]

arXiv preprint arXiv:2401.11673 , year=

Cao,C.,Ren,X.,Fu,Y.:Mvsformer++:Revealingthedevilintransformer’sdetails for multi-view stereo. arXiv preprint arXiv:2401.11673 (2024)

work page arXiv 2024
[4]

In: Proc

Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)

2017
[5]

In: Proceedings of the IEEE international conference on computer vision (ICCV)

Dosovitskiy, A., Fischer, P., Ilg, E., Hausser, P., Hazirbas, C., Golkov, V., Van Der Smagt, P., Cremers, D., Brox, T.: Flownet: Learning optical flow with convolu- tional networks. In: Proceedings of the IEEE international conference on computer vision (ICCV). pp. 2758–2766 (2015)

2015
[6]

In: Conference on Computer Vision and Pattern Recogni- tion (CVPR) (2012)

Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recogni- tion (CVPR) (2012)

2012
[7]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset generator. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3749–3761 (2022)

2022
[8]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: 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. pp. 2495–2504 (2020)

2020
[9]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

Hu, Y.T., Wang, J., Huang, J.B., Schwing, A.G.: Sail-vos 3d: A video dataset for self-supervised 3d understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

2021
[10]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

2018
[11]

In: Proceedings of the Computer Vision and Pattern Recognition Conference

Izquierdo, S., Sayed, M., Firman, M., Garcia-Hernando, G., Turmukhambetov, D., Civera, J., Mac Aodha, O., Brostow, G., Watson, J.: Mvsanywhere: Zero-shot multi-view stereo. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 11493–11504 (2025) 16 Zeyu Ma, Alexander Raistrick, and Jia Deng

2025
[12]

In: Proceedings of the Computer Vision and Pattern Recognition Conference

Jiang, H., Xu, Z., Xie, D., Chen, Z., Jin, H., Luan, F., Shu, Z., Zhang, K., Bi, S., Sun, X., et al.: Megasynth: Scaling up 3d scene reconstruction with synthesized data. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16441–16452 (2025)

2025
[13]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)

Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Dynamicstereo: Consistent dynamic depth from stereo videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)

2023
[14]

In: ACM Transactions on Graphics (ToG) (2017)

Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. In: ACM Transactions on Graphics (ToG) (2017)

2017
[15]

ACM Transactions on Graphics36(4) (2017)

Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics36(4) (2017)

2017
[16]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)

Li, Y., Jiang, L., Xu, L., Xiangli, Y., Wang, Z., Lin, D., Dai, B.: Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)

2023
[17]

In: Eu- ropean Conference on Computer Vision

Ma, Z., Teed, Z., Deng, J.: Multiview stereo with cascaded epipolar raft. In: Eu- ropean Conference on Computer Vision. pp. 734–750. Springer (2022)

2022
[18]

In: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR)

Mayer, N., Ilg, E., Hausser, P., Fischer, P., Fischer, C., Cremers, D., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR). pp. 4040–4048 (2016)

2016
[19]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Raistrick, A., Mei, L., Kayan, K., Yan, D., Zuo, Y., Han, B., Wen, H., Parakh, M., Alexandropoulos, S., Lipson, L., et al.: Infinigen indoors: Photorealistic indoor scenes using procedural generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21783–21794 (2024)

2024
[20]

CVPR (2023)

Raistrick, A., Zhai, C., Ma, Z., Mei, L., Wang, Y., Yi, K., Sun, W., Ho, C.H., Wang, C., Wang, J., et al.: Infinigen: Infinite photorealistic worlds using procedural generation. CVPR (2023)

2023
[21]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)

Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Angelova, A., Applehoff, N., Bautista, M.A.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)

2021
[22]

In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)

Schöps, T., Schönberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)

2017
[23]

In: 2022 International Conference on 3D Vision (3DV)

Schröppel, P., Bechtold, J., Amiranashvili, A., Brox, T.: A benchmark and a base- line for robust multi-view depth estimation. In: 2022 International Conference on 3D Vision (3DV). pp. 406–415 (2022).https://doi.org/10.1109/3DV57658. 2022.00052,https://arxiv.org/abs/2209.06681

work page doi:10.1109/3dv57658 2022
[24]

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

Shi, X., Huang, Z., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Flowformer++: Masked cost volume autoencoding for pretraining optical flow estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1599–1610 (2023)

2023
[25]

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

Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatch- net: Learned multi-view stereo with deep patchmatch. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4414–4424 (2021)

2021
[26]

In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2020) SimpleProc: Fully Procedural Synthetic Data from Simple Rules for MVS 17

Wang, W., Zhu, D., Wang, X., Hu, Y., Qiu, Y., Wang, C., Hu, Y., Kapoor, A., Scherer, S.: Tartanair: A dataset to push the limits of visual slam. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2020) SimpleProc: Fully Procedural Synthetic Data from Simple Rules for MVS 17

2020
[27]

W AFT: Warping-Alone Field Transforms for Optical Flow.arXiv preprint arXiv:2506.21526, 2025

Wang, Y., Deng, J.: Waft: Warping-alone field transforms for optical flow. arXiv preprint arXiv:2506.21526 (2025)

work page arXiv 2025
[28]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Yang, D., Deng,J.: Shape fromshadingthrough shape evolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3781–3790 (2018)

2018
[29]

Depth Anything V2

Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything v2. arXiv preprint arXiv:2406.09414 (2024)

work page internal anchor Pith review arXiv 2024
[30]

In: Proceedings of the European conference on computer vision (ECCV)

Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstruc- tured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767–783 (2018)

2018
[31]

In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

Yao, Y., Luo, Z., Li, S., Zhang, J., Ren, S., Zhou, L., Fang, T., Quan, L.: Blended- mvs: A large-scale dataset for generalized multi-view stereo networks. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

2020