pith. sign in

arxiv: 1906.11557 · v1 · pith:NSTWPNT3new · submitted 2019-06-27 · 💻 cs.GR

Flexible SVBRDF Capture with a Multi-Image Deep Network

Valentin Deschaintre , Miika Aittala , Fredo Durand , George Drettakis , Adrien Bousseau This is my paper

Pith reviewed 2026-05-25 14:06 UTC · model grok-4.3

classification 💻 cs.GR
keywords SVBRDFmaterial capturedeep learningmulti-imagereflectance estimationhandheld camerauncalibrated imagesfusing layer
0
0 comments X

The pith

A deep network estimates material reflectance from any number of uncalibrated handheld photos.

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

The paper introduces a neural network architecture that accepts a variable number of photos taken with a consumer camera and flash, without calibration or ordering of the inputs, and produces an estimate of the material's spatially varying reflectance properties. It improves reconstruction quality as more images are added, reaching good results with between one and ten pictures. The approach relies on learned priors from training data combined with a special layer that fuses per-image information independently of sequence. This sits between single-image methods that lack detail and traditional multi-image methods that demand careful setup.

Core claim

The authors present a deep-learning method that estimates SVBRDF appearance from a variable number of uncalibrated and unordered pictures captured with a handheld camera and flash, using an order-independent fusing layer to extract useful information from each input while benefiting from data-driven priors, and handling both view and light direction variation without calibration.

What carries the argument

Order-independent fusing layer that combines information from each input image regardless of order or calibration status.

If this is right

  • Reconstruction quality increases as the number of input pictures grows.
  • High-quality results are possible with as few as one to ten images.
  • The method works on uncalibrated and unordered inputs from handheld capture.
  • View and light direction changes are handled without separate calibration steps.

Where Pith is reading between the lines

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

  • Consumer devices could capture usable material data in everyday uncontrolled settings.
  • The same fusion approach might apply to other inverse problems that receive variable numbers of observations.
  • Performance on materials outside the training distribution remains an open test case.

Load-bearing premise

A network trained on data with learned priors can generalize to real uncalibrated photos and reliably extract useful per-image information without explicit calibration or ordering.

What would settle it

Capture a set of real handheld photos of a material with known ground-truth SVBRDF parameters, feed them to the network, and check whether the output parameters match the ground truth within measurement error.

Figures

Figures reproduced from arXiv: 1906.11557 by Adrien Bousseau, Fredo Durand, George Drettakis, Miika Aittala, Valentin Deschaintre.

Figure 1
Figure 1. Figure 1: Our deep learning method for SVBRDF capture supports a variable number of input photographs taken with uncalibrated light￾view directions (a, rectified). While a single image is enough to obtain a first plausible estimate of the SVBRDF maps, more images provide new cues to our method, improving its prediction. In this example, adding images reveals fine normal variations (b), removes highlight residuals in… view at source ↗
Figure 2
Figure 2. Figure 2: We use a simple paper frame to help register pictures taken from different viewpoints. We use either a single smartphone and its flash, or two smartphones to cover a larger set of view/light configurations. 64-channel feature map, which retains more information to be pro￾cessed by the later stages of our architecture. We also provide pixel coordinates as extra channels to the input to help the convolutiona… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our deep network architecture. Each input image is processed by its copy of the encoder-decoder to produce a feature map. While the number of images and network copies can vary, a pooling layer fuses the output maps to obtain a fixed-size representation of the material, which is then processed by a few convolutional layers to produce the SVBRDF maps. 4.4. Training We train our network for 7 day… view at source ↗
Figure 4
Figure 4. Figure 4: SSIM of our predictions with respect to the number of input images, averaged over our synthetic test dataset. The SSIM of re-renderings increases quickly for the first images, before stabi￾lizing at around 10 images. The normal maps strongly benefit from new images. Diffuse and specular albedos also improve with addi￾tional inputs, which is not the case of the roughness that remains stable overall. We prov… view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study. Comparison of SSIM between our method (green) and a restricted version (black) where the network is trained with lighting and viewing directions chosen on a perfect hemisphere, and with all lighting parameters constant (falloff ex￾ponent, power, etc.). Our complete method achieves higher SSIM when tested on a dataset with small variations of these parameters, showing that it is robust to su… view at source ↗
Figure 6
Figure 6. Figure 6: Evaluation on a measured BTF. Three images are enough to capture most of normal and roughness maps. Adding images further improves the result by removing lighting residual from the diffuse albedo, and adding subtle details to the normal and specular maps. 7. Conclusion With the advance of deep learning, the holy grail of single-image SVBRDF capture recently became a reality. Yet, despite impressive results… view at source ↗
Figure 7
Figure 7. Figure 7: A single flash picture hardly provides enough information for surfaces composed of several materials. In this example, adding images allows the recovery of normal details, and the capture of different roughness values in different parts of the image. Note in particular how the 4th image helps capturing a discontinuity of the roughness on the right part. [CHW18] CHEN G., HAN K., WONG K.-Y. K.: Ps-fcn: A fle… view at source ↗
Figure 8
Figure 8. Figure 8: SSIM on re-renderings for the maps obtained by our method with 5 images (dotted blue) and by a classical optimiza￾tion method with an increasing number of input images (black). The classical optimization requires several dozens of calibrated pictures to outperform our method on rather diffuse or uniform ma￾terials (stones, tiles, scales), while requiring many more for a more complex material (wood). pearan… view at source ↗
Figure 9
Figure 9. Figure 9: Comparison against single-image methods on synthetic SVBRDFs. Our method leverages additional input images to obtain SVBRDF maps closer to ground truth. In particular, single-image methods under-estimate normal variations and fail to remove the sat￾urated highlight on shiny materials. See supplemental materials for more comparisons and results. 2019 Authors version [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison against single-image methods on real-world pictures. Our method recovers more normal details, and better removes highlight and shading residuals from the diffuse albedo. See supplemental materials for more comparisons and results. 2019 Authors version [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison to real-world relighting. Each column shows re-renderings of a captured material, except the last column which shows a picture of that material under a similar lighting con￾dition (not used as input). We manually adjusted the position of the virtual light to best match the ground truth. Similarly, we adjusted the light power for each method separately since each has its own arbitrary scale fact… view at source ↗
Figure 12
Figure 12. Figure 12: Comparison against single-image methods on a measured BTF with ground truth re-renderings. Our method globally captures the material features better. 2019 Authors version [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Limitations. We inherits some of the limitations of the method by Deschaintre et al. [DAD∗ 18], such as the tendency to produce correlated maps and to interpret dark pixels as shiny (top). Our SVBRDF representation, training data and loss do not model cast shadows. As a result, shadows in the input pollute some of the maps (bottom). 2019 Authors version [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
read the original abstract

Empowered by deep learning, recent methods for material capture can estimate a spatially-varying reflectance from a single photograph. Such lightweight capture is in stark contrast with the tens or hundreds of pictures required by traditional optimization-based approaches. However, a single image is often simply not enough to observe the rich appearance of real-world materials. We present a deep-learning method capable of estimating material appearance from a variable number of uncalibrated and unordered pictures captured with a handheld camera and flash. Thanks to an order-independent fusing layer, this architecture extracts the most useful information from each picture, while benefiting from strong priors learned from data. The method can handle both view and light direction variation without calibration. We show how our method improves its prediction with the number of input pictures, and reaches high quality reconstructions with as little as 1 to 10 images -- a sweet spot between existing single-image and complex multi-image approaches.

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.

Referee Report

3 major / 2 minor

Summary. The paper presents a deep neural network for estimating SVBRDFs from a variable number of uncalibrated, unordered images captured with a handheld camera and flash. The core contribution is an order-independent fusing layer that aggregates per-image features while benefiting from data-driven priors, enabling the method to handle view/light variation without calibration and to improve output quality as the number of inputs increases from 1 to ~10.

Significance. If the empirical results hold, the work would provide a practical middle ground between single-image DL methods and traditional multi-image optimization, lowering the barrier to high-quality material acquisition using consumer hardware. The flexible input cardinality and learned priors are potentially impactful for graphics pipelines if the generalization from training data to real handheld captures is demonstrated.

major comments (3)
  1. [Abstract and §5] Abstract and §5 (Results): the central performance claims ('improves its prediction with the number of input pictures' and 'reaches high quality reconstructions with as little as 1 to 10 images') are stated without accompanying quantitative metrics, error bars, baseline comparisons, or cross-validation on real captures. This leaves the generalization claim (implicit calibration via the fusing layer) unsupported in the provided text.
  2. [§4.2] §4.2 (Order-independent fusing layer): the layer is described as extracting 'the most useful information from each picture' without explicit calibration, yet no ablation is reported that isolates its contribution versus standard pooling or concatenation. Without this, it is unclear whether the layer is load-bearing for the variable-input claim or whether simpler architectures would suffice.
  3. [§3 and §6] §3 and §6 (Training data and real-world evaluation): the method relies on strong priors learned from (presumably synthetic) data to disambiguate lighting/view directions on real uncalibrated inputs. The manuscript does not detail the training distribution statistics or provide failure-case analysis on real sensor noise, flash falloff, or pose distributions that diverge from training.
minor comments (2)
  1. [§4] Notation for the fusing layer (Eq. in §4) could be clarified with a small diagram or pseudocode showing how order invariance is enforced while preserving per-image lighting cues.
  2. [Figure 1 and §5] Figure 1 and §5 examples would benefit from explicit captions stating the exact number of input images and whether they are synthetic or real.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below and commit to revisions that strengthen the empirical support and clarity of the manuscript.

read point-by-point responses

  1. Referee: [Abstract and §5] Abstract and §5 (Results): the central performance claims ('improves its prediction with the number of input pictures' and 'reaches high quality reconstructions with as little as 1 to 10 images') are stated without accompanying quantitative metrics, error bars, baseline comparisons, or cross-validation on real captures. This leaves the generalization claim (implicit calibration via the fusing layer) unsupported in the provided text.

    Authors: We agree that the current text would benefit from quantitative backing. In the revision we will add error metrics (with standard deviations) on a held-out synthetic test set for 1–10 inputs, direct comparisons to single-image baselines and multi-image optimization, and additional real-world visual results. We will note that pixel-accurate ground truth is unavailable for real handheld captures and therefore rely on visual assessment there. revision: yes

  2. Referee: [§4.2] §4.2 (Order-independent fusing layer): the layer is described as extracting 'the most useful information from each picture' without explicit calibration, yet no ablation is reported that isolates its contribution versus standard pooling or concatenation. Without this, it is unclear whether the layer is load-bearing for the variable-input claim or whether simpler architectures would suffice.

    Authors: We will add an ablation study comparing the order-independent fusing layer to mean/max pooling and concatenation baselines, reporting performance across varying input cardinalities to isolate its contribution to the variable-input and order-independent behavior. revision: yes

  3. Referee: [§3 and §6] §3 and §6 (Training data and real-world evaluation): the method relies on strong priors learned from (presumably synthetic) data to disambiguate lighting/view directions on real uncalibrated inputs. The manuscript does not detail the training distribution statistics or provide failure-case analysis on real sensor noise, flash falloff, or pose distributions that diverge from training.

    Authors: We will expand §3 with explicit statistics on the synthetic training distribution (material parameters, lighting/view ranges). We will also add a limitations subsection in §6 that discusses and illustrates failure modes arising from sensor noise, flash falloff, and pose distributions outside the training support. revision: yes

Circularity Check

0 steps flagged

No circularity; trained network architecture with empirical generalization

full rationale

The paper describes a deep network with an order-independent fusing layer trained on synthetic data to map variable uncalibrated images to SVBRDF parameters. No equations, derivations, or self-citations are present in the provided text that reduce any claimed prediction to its inputs by construction, rename a fit as a prediction, or import uniqueness via author citations. The method's correctness is framed as empirical performance on real images, which is independent of the listed circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The claim rests on the generalization ability of a data-driven network and the effectiveness of the introduced fusing layer; both are learned components without independent verification in the provided abstract.

free parameters (1)
  • network weights
    Deep network parameters fitted during training on material appearance data.
axioms (1)
  • domain assumption Training data distribution is representative of real-world materials under varying view and light conditions
    The method depends on learned priors to handle uncalibrated inputs.
invented entities (1)
  • order-independent fusing layer no independent evidence
    purpose: Combine information from variable numbers of unordered images
    New architectural component introduced to enable the multi-image capability.

pith-pipeline@v0.9.0 · 5693 in / 1350 out tokens · 31959 ms · 2026-05-25T14:06:36.571510+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Rendering-Aware Sparse Sampling for BRDF Acquisition

    cs.CV 2026-04 unverdicted novelty 6.0

    A sampler network learns to select informative sparse BRDF measurement directions by optimizing against a fixed pretrained hypernetwork reconstructor and differentiable renderer, improving low-budget reconstruction on...

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

    work page

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry.original add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 i...

    work page

  3. [3]

    Abadi M., Agarwal A., Barham P., Brevdo E., Chen Z., Citro C., Corrado G. S., Davis A., Dean J., Devin M., Ghemawat S., Goodfellow I., Harp A., Irving G., Isard M., Jia Y., Jozefowicz R., Kaiser L., Kudlur M., Levenberg J., Man\' e D., Monga R., Moore S., Murray D., Olah C., Schuster M., Shlens J., Steiner B., Sutskever I., Talwar K., Tucker P., Vanhoucke...

    work page 2015

  4. [4]

    : Reflectance modeling by neural texture synthesis

    Aittala M., Aila T., Lehtinen J. : Reflectance modeling by neural texture synthesis. ACM Transactions on Graphics (Proc. SIGGRAPH) 35, 4 (2016)

    work page 2016

  5. [5]

    : Burst image deblurring using permutation invariant convolutional neural networks

    Aittala M., Durand F. : Burst image deblurring using permutation invariant convolutional neural networks. In The European Conference on Computer Vision (ECCV) (2018)

    work page 2018

  6. [6]

    URL: https://share.allegorithmic.com/

    Allegorithmic : Substance share, 2018. URL: https://share.allegorithmic.com/

    work page 2018

  7. [7]

    : Practical SVBRDF capture in the frequency domain

    Aittala M., Weyrich T., Lehtinen J. : Practical SVBRDF capture in the frequency domain

    work page

  8. [8]

    : Two-shot SVBRDF capture for stationary materials

    Aittala M., Weyrich T., Lehtinen J. : Two-shot SVBRDF capture for stationary materials. ACM Trans. Graph. (Proc. SIGGRAPH) 34, 4 (July 2015), 110:1--110:13. URL: http://doi.acm.org/10.1145/2766967, https://doi.org/10.1145/2766967 doi:10.1145/2766967

    work page doi:10.1145/2766967 2015

  9. [9]

    Chen G., Han K., Wong K.-Y. K. : Ps-fcn: A flexible learning framework for photometric stereo. In The European Conference on Computer Vision (ECCV) (2018)

    work page 2018

  10. [10]

    L., Torrance K

    Cook R. L., Torrance K. E. : A reflectance model for computer graphics. ACM Transactions on Graphics 1, 1 (1982), 7--24

    work page 1982

  11. [11]

    B., Xu D., Gwak J., Chen K., Savarese S

    Choy C. B., Xu D., Gwak J., Chen K., Savarese S. : 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In IEEE European Conference on Computer Vision (ECCV) (2016), pp. 628--644

    work page 2016

  12. [12]

    : Single-image svbrdf capture with a rendering-aware deep network

    Deschaintre V., Aittala M., Durand F., Drettakis G., Bousseau A. : Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (SIGGRAPH Conference Proceedings) 37, 128 (aug 2018), 15. URL: http://www-sop.inria.fr/reves/Basilic/2018/DADDB18

    work page 2018

  13. [13]

    : Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting

    Dong Y., Chen G., Peers P., Zhang J., Tong X. : Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics (Proc. SIGGRAPH Asia) 33, 6 (2014)

    work page 2014

  14. [14]

    J., Van Ginneken B., Nayar S

    Dana K. J., Van Ginneken B., Nayar S. K., Koenderink J. J. : Reflectance and texture of real-world surfaces. ACM Transactions On Graphics (TOG) 18, 1 (1999), 1--34

    work page 1999

  15. [15]

    : Manifold bootstrapping for svbrdf capture

    Dong Y., Wang J., Tong X., Snyder J., Ben-Ezra M., Lan Y., Guo B. : Manifold bootstrapping for svbrdf capture. ACM Transactions on Graphics (Proc. SIGGRAPH) 29, 4 (2010)

    work page 2010

  16. [16]

    A., Debevec P

    Ghosh A., Chen T., Peers P., Wilson C. A., Debevec P. : Estimating specular roughness and anisotropy from second order spherical gradient illumination. In Computer Graphics Forum (June 2009), vol. 28, p. 4

    work page 2009

  17. [17]

    C., Ghosh A., Denk C., Glencross M

    Guarnera D., Guarnera G. C., Ghosh A., Denk C., Glencross M. : BRDF Representation and Acquisition . Computer Graphics Forum (2016)

    work page 2016

  18. [18]

    : Linear light source reflectometry

    Gardner A., Tchou C., Hawkins T., Debevec P. : Linear light source reflectometry. ACM Trans. Graph. 22, 3 (July 2003), 749--758. URL: http://doi.acm.org/10.1145/882262.882342, https://doi.org/10.1145/882262.882342 doi:10.1145/882262.882342

    work page doi:10.1145/882262.882342 2003

  19. [19]

    Y., Hadap S., Wang J., Sankaranarayanan A

    Hui Z., Sunkavalli K., Lee J. Y., Hadap S., Wang J., Sankaranarayanan A. C. : Reflectance capture using univariate sampling of brdfs. In IEEE International Conference on Computer Vision (ICCV) (2017)

    work page 2017

  20. [20]

    P., Ba J

    Kingma D. P., Ba J. : Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR) (2015)

    work page 2015

  21. [21]

    : Efficient reflectance capture using an autoencoder

    Kang K., Chen Z., Wang J., Zhou K., Wu H. : Efficient reflectance capture using an autoencoder. ACM Transactions on Graphics (Proc. SIGGRAPH) 37, 4 (July 2018)

    work page 2018

  22. [22]

    : Material editing using a physically based rendering network

    Liu G., Ceylan D., Yumer E., Yang J., Lien J.-M. : Material editing using a physically based rendering network. In IEEE International Conference on Computer Vision (ICCV) (2017), pp. 2261--2269

    work page 2017

  23. [23]

    : Modeling surface appearance from a single photograph using self-augmented convolutional neural networks

    Li X., Dong Y., Peers P., Tong X. : Modeling surface appearance from a single photograph using self-augmented convolutional neural networks. ACM Transactions on Graphics (Proc. SIGGRAPH) 36, 4 (2017)

    work page 2017

  24. [24]

    An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution

    Liu R., Lehman J., Molino P., Such F. P., Frank E., Sergeev A., Yosinski J. : An intriguing failing of convolutional neural networks and the coordconv solution. CoRR abs/1807.03247 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  25. [25]

    : Reflectance and illumination recovery in the wild

    Lombardi S., Nishino K. : Reflectance and illumination recovery in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) 38 (2016), 129--141

    work page 2016

  26. [26]

    : Materials for masses: SVBRDF acquisition with a single mobile phone image

    Li Z., Sunkavalli K., Chandraker M. : Materials for masses: SVBRDF acquisition with a single mobile phone image. Proceedings of ECCV (2018)

    work page 2018

  27. [27]

    : Learning to reconstruct shape and spatially-varying reflectance from a single image

    Li Z., Xu Z., Ramamoorthi R., Sunkavalli K., Chandraker M. : Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (Proc. SIGGRAPH Asia) (2018)

    work page 2018

  28. [28]

    Mcallister D. K. : A Generalized Surface Appearance Representation for Computer Graphics. PhD thesis, 2002

    work page 2002

  29. [29]

    A., Claus D., Fitzgibbon A

    Paterson J. A., Claus D., Fitzgibbon A. W. : Brdf and geometry capture from extended inhomogeneous samples using flash photography. Computer Graphics Forum (Proc. Eurographics) 24, 3 (Sept. 2005), 383--391

    work page 2005

  30. [30]

    R., Su H., Mo K., Guibas L

    Qi C. R., Su H., Mo K., Guibas L. J. : Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)

    work page 2017

  31. [31]

    V., Tuytelaars T

    Rematas K., Georgoulis S., Ritschel T., Gavves E., Fritz M., Gool L. V., Tuytelaars T. : Reflectance and natural illumination from single-material specular objects using deep learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2017)

    work page 2017

  32. [32]

    : U-net: Convolutional networks for biomedical image segmentation

    Ronneberger O., P.Fischer, Brox T. : U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2015), vol. 9351 of LNCS, pp. 234--241

    work page 2015

  33. [33]

    : Mobile surface reflectometry

    Riviere J., Peers P., Ghosh A. : Mobile surface reflectometry. Computer Graphics Forum 35, 1 (2016)

    work page 2016

  34. [34]

    : Polarization imaging reflectometry in the wild

    Riviere J., Reshetouski I., Filipi L., Ghosh A. : Polarization imaging reflectometry in the wild. ACM Transactions on Graphics (Proc. SIGGRAPH) (2017)

    work page 2017

  35. [35]

    : Pocket reflectometry

    Ren P., Wang J., Snyder J., Tong X., Guo B. : Pocket reflectometry. ACM Transactions on Graphics (Proc. SIGGRAPH) 30, 4 (2011)

    work page 2011

  36. [36]

    : Material classification based on training data synthesized using a btf database

    Weinmann M., Gall J., Klein R. : Material classification based on training data synthesized using a btf database. In European Conference on Computer Vision (ECCV) (2014), pp. 156--171

    work page 2014

  37. [37]

    : Estimating dual-scale properties of glossy surfaces from step-edge lighting

    Wang C.-P., Snavely N., Marschner S. : Estimating dual-scale properties of glossy surfaces from step-edge lighting. ACM Transactions on Graphics (Proc. SIGGRAPH Asia) 30, 6 (2011)

    work page 2011

  38. [38]

    : Silnet : Single- and multi-view reconstruction by learning from silhouettes

    Wiles O., Zisserman A. : Silnet : Single- and multi-view reconstruction by learning from silhouettes. British Machine Vision Conference (BMVC) (2017)

    work page 2017

  39. [39]

    : Single image surface appearance modeling with self-augmented cnns and inexact supervision

    Ye W., Li X., Dong Y., Peers P., Tong X. : Single image surface appearance modeling with self-augmented cnns and inexact supervision. Computer Graphics Forum 37, 7 (2018), 201--211

    work page 2018

  40. [40]

    R., Smola A

    Zaheer M., Kottur S., Ravanbakhsh S., Poczos B., Salakhutdinov R. R., Smola A. J. : Deep sets. In Advances in Neural Information Processing Systems (NIPS). 2017

    work page 2017