learning-based rendering - kaist
TRANSCRIPT
Learning-Based Rendering
2019.05.16.
20193138 Jaeyoon Kim
1
Review for previous presentation
β’ βMicrofacet Model and Material Appearanceβ by Hakyeong Kim
2
Machine Learning in Rendering
From CheolMinβs slides From Saehunβs slides3
Deep Scattering:Rendering Atmospheric Clouds withRadiance-Predicting Neural Networks
by S. Kallweit, T. MΓΌller, B. McWilliams, M. Gross, J. NovΓ‘k
Slides and images from βDeep Learning for Light Transportβ by Jan NovΓ‘k, Disney Research 4
Remind previous presentation
From Hyunsooβs slides 5
6
7
8
9
Rendering Clouds
πΏπΏ π±π±,π€π€ = οΏ½0
βππ π²π²|π±π± πΌπΌοΏ½
ππππ οΏ½π€π€|π€π€ πΏπΏ π²π², οΏ½π€π€ dοΏ½π€π€dπ¦π¦
Predict using a network
10
Rendering Clouds
β’ Scattered direct illuminationβ’ Estimated using MC
11
Rendering Clouds
β’ Scattered direct illuminationβ’ Estimated using MC
β’ Scattered indirect illuminationβ’ Predicted using NN
12
Radiance Predicting MLP
β’ Uses hierarchical stencil for sampling.β’ for both local details and overall shape
β’ Feeds into the network.
13
Radiance Predicting MLP
β’ Uses hierarchical stencil for sampling.β’ for both local details and overall shape
β’ Feeds into the network.β’ One block with two layers for each stencil.β’ Some FC layers at the end of the network.
14
15
16
Slides and images from βBayesian online regression for adaptive direct illumination
samplingβ by Petr at al. 17
Noise in MC Rendering
18
Noise in MC Rendering
Non-adaptive sampling[Wang et al. 2009]
Reference
19
Noise in MC Rendering
Non-adaptive sampling[Wang et al. 2009]
Adaptive sampling[Donikian et al. 2006]
20
Direct Illumination
Less important
Occluded
21
Light Clustering
22
Direct Illumination
β’ πΏπΏ π±π±,π€π€ = β«π΄π΄ πΉπΉ π²π² β π±π± β π€π€ dπ²π²
β’ πΉπΉ π²π² β π±π± β π€π€ = πΏπΏππ π²π² β π±π± π΅π΅ π²π² β π±π± β π€π€ ππ π²π² β π±π± πΊπΊ π²π² β π±π±
β’ πΏπΏ π±π±,π€π€ = πΉπΉ π²π²βπ±π±βπ€π€ππ π²π²|π±π±,π€π€
π€π€
π²π² π±π±π΄π΄
23
Probability Density Function
β’ ππ π²π²|π±π±,π€π€ = ππ ππ|π±π± ππ ππ|ππ ππ π²π²|ππ,π€π€β’ ππ ππ|π±π± : selecting a light cluster c β Cβ’ ππ ππ|ππ : selecting a light ππ β ππ, ππ ππ|ππ = π·π·ππ/βππββππ π·π·ππβ
β’ ππ π²π²|ππ,π€π€ : selecting a point π¦π¦ β ππ, using standard techniques in [Pharr et al. 2016; Shirley et al. 1996].
β’ Only need to predict ππ ππ|π±π± .
24
Optimal Cluster Selection
β’ ππππππ πΏπΏ π±π± = βπΏπΏ π±π± 2 + βππβπΆπΆ1
ππ ππ|π±π± β«π΄π΄πππΉπΉ π²π²βπ±π± 2
ππ ππ|ππ ππ π²π²|ππdπ²π²
β’ ππ2,ππ = β«π΄π΄πππΉπΉ π²π²βπ±π± 2
ππ ππ|ππ ππ π²π²|ππdπ²π²: second moment of πΏπΏππ π±π± = πΉπΉ π²π²βπ±π±
ππ ππ|ππ ππ π²π²|ππ
β’ ππππππππ ππ|π±π± β πΏπΏππ2 π±π± + ππππππ πΏπΏππ π±π±
25
Bayesian Online Regression
β’ ππ: training dataβ’ ππ: parametersβ’ ππ ππ|ππ : a model describing the likelihood of ππ given ππ
β’ Maximum likelihood (ML) estimateβ’ Finding ππ maximizing ππ ππ|ππβ’ Prone to overfit when data is scarceβ’ Gives poor approximations in early stages of rendering
β’ Maximum a posteriori (MAP) estimateβ’ Finding ππ maximizing ππ ππ|ππ β ππ ππ|ππ ππ ππβ’ More robust than ML estimate
26
Bayesian Online Regression
β’ οΏ½ΜοΏ½π = πΏπΏππ π²π²βπ±π± ππ π²π²βπ±π± cosπππ²π²ππ2 π±π±,π²π² ππ ππ|ππ ππ π²π²|ππ
, οΏ½ΜοΏ½ππ±π± = οΏ½ΜοΏ½πcosπππ±π±
β’ Data ππ consists of tuples οΏ½ΜοΏ½ππ±π±,ππ , οΏ½ΜοΏ½πππ .
β’ ππ(ππ|ππ) = βππππ ππ οΏ½ΜοΏ½ππ±π±,ππ|οΏ½ΜοΏ½πππ ,ππ
β’ ππ οΏ½ΜοΏ½ππ±π±|οΏ½ΜοΏ½π,ππ = πΏπΏ οΏ½ΜοΏ½ππ±π± ππ0 + 1 β ππ0 ππ οΏ½ΜοΏ½ππ±π±| πποΏ½ππ2
, βοΏ½ππ4β’ ππ = ππ0, ππ, ββ’ ππ ππ = π΅π΅ ππ0| οΏ½ππππ, οΏ½πππ£π£ ππ β π€π€β1 ππ, β|ππ0, οΏ½ππ, οΏ½πππΌπΌ ,π½π½
27
Bayesian Online Regression
β’ πΏπΏππ π±π± β 1βππ0 πποΏ½ππ2
β’ ππππππ πΏπΏππ π±π± β 1βππ0 ππ0ππ2+βοΏ½ππ4
β’ ππβ ππ|π±π± β 1οΏ½ππ2
1 β ππ0 2ππ2 + 1 β ππ0 ππ0ππ2 + β
28
Summary
β’ Light preprocess (clustering)β’ During each Next event estimation:
β’ Obtain cached clustering.β’ Compute distributions of estimates for each cluster. (mean, variance)β’ Build distribution over clusters.β’ Sample direct illumination.β’ Record new data for sampled cluster.
29
Results
30
Results
31
Reference
β’ Simon Kallweit, Thomas Muller, Brian McWilliams, Markus Gross, and Jan Novak. 2017. Deep Scattering: Rendering Atmospheric Clouds with Radiance-Predicting Neural Networks. ACM Trans. Graph. 36, 6, Article 231 (November 2017), 11 pages.
β’ J. NovΓ‘k at al. 2018. Deep Learning for Light Transport. Disney Research.
β’ Petr Vevoda, Ivo Kondapaneni, and Jaroslav KΕivΓ‘nek. 2018. Bayesian online regression for adaptive direct illumination sampling. ACM Trans. Graph. 37, 4, Article 125 (August 2018), 12 pages.
32
Thank you!Any questions?
33
Quiz
Fill in the blanks with words in the box.
1. In βDeep Scattering β¦β paper, scattered ( ) illumination is predicted using a neural network.
2. In βBayesian online regression β¦β paper, authors aim to reduce noise in ( ) illumination with Bayesian online learning.
(a) Direct(b) Indirect(c) Diffuse(d) Specular
34