進階搜尋


下載電子全文  
系統識別號 U0026-3012201923245000
論文名稱(中文) SenroDR - 新的可微分渲染框架
論文名稱(英文) SenroDR - A New Differentiable Rendering Framework
校院名稱 成功大學
系所名稱(中) 資訊工程學系
系所名稱(英) Institute of Computer Science and Information Engineering
學年度 108
學期 1
出版年 108
研究生(中文) 許友綸
研究生(英文) Yu-Lun Hsu
學號 P76082015
學位類別 碩士
語文別 英文
論文頁數 53頁
口試委員 指導教授-朱威達
共同指導教授-胡敏君
口試委員-蔡侑庭
口試委員-朱宏國
口試委員-陳華總
中文關鍵字 光跡跟蹤  渲染  微分 
英文關鍵字 Differential  Rendering  Path Tracing 
學科別分類
中文摘要 渲染意旨將數位資訊轉化為圖片或影片媒體的過程。由於場景的不連續性,長久以來渲染被認為不可能微分。可微分渲染今天仍然是一個相對較新的主題,2018年才由李子懋 [1]首先找出其解。現在市面上有兩個通用可微分渲染器(以下簡稱DR),即OpenDR [2]與Redner,當然還有其他特殊用途或是玩具實作 [3] [4]。然而,這些現有的DR都無法滿足我們的需求。我們想要一個快速,且合理使用記憶體,並且具有彈性的DR供我們做研究。因此,我們做了一個新的DR - SenroDR。我們也將在此介紹SenroDR框架。本篇論文也將redner的邊緣採樣法擴展為曲線採樣法,這使我們支援更多不同的物體形狀,而我們也在SenroDR支援二次曲面與其微分。而實驗將以展示本論文contribution為主軸去設計,檢驗本研究之有效性、效能(速度和記憶體用量)、曲面採樣的正確性、彈性
英文摘要 Rendering is the process of converting digital information into image or video media. Due to the discontinuity of the scene, rendering has long been considered impossible to differentiate. Differetiable rendering(DR) is a relatively new topic in computer graphics. It’s unbiased solution was found in the year 2018 by Lee [1]. There are two existing general-purpose DR in the world and some special-purpose or toy DR available. However, None of them meet our need. We need a more efficient in both speed and memory consumption and flexible DR for research. Thus, here we present our new differentiable renderer, SenroDR. In thesis, we briefly introduce and break down our SenroDR framework. Our other contribution is mathematical derivation of curve sampling, which is extended from edge sampling technique developed by Lee [1].And we designed several different experiments based on the contribution of the paper.
論文目次 中文摘要 i
Abstract ii
Acknowledgements iii
Contents iv
List of Figures vii
1 Introduction 1
1.1 Related Work 2
1.1.1 Inverse Rendering 3
1.1.2 Differentiable Rendering 3
1.2 Contribution 4
1.3 Thesis Organization 4
1.4 Mathematical Notation 5
2 Understand Rendering 6
2.1 Light Field 6
2.2 Rendering Equation 7
2.3 Monte Carlo Integration8
2.3.1 Importance Sampling 8
2.4 Path Tracing 9
2.5 Measurement Interpretation 11
3 Differentiable Rendering 12
3.1 OpenDR 13
3.1.1 Forward Process 13
3.1.2 Backward Process 14
3.1.3 Differentiating Appearance 14
3.1.4 Differentiating Projection 15
3.1.5 Calculate Gradient w.r.t Desired Variable 15
3.2 Mathematical Formulation 15
3.2.1 The First Integral 17
3.2.2 The Second Integral 19
3.3 Problem of OpenDR 19
3.4 Secondary Visibility 20
3.5 Implementation Hint 20
3.5.1 Autodiff/Autograd 20
3.6 Toy Example 21
3.7 Curve Sampling 23
4 SenroDR Framework 24
4.1 Senro 24
4.2 System Framework 24
4.3 Scene Arrangement 25
4.3.1 Pytorch Autograd Function 26
4.4 Camera Setupp 26
4.5 Closest Intersect Calculation 27
4.5.1 Memory Usage Reduction 29
4.6 Active Ray Filtering 31
4.7 Hit Record Calculation 31
4.8 Rays Generation 32
4.9 Shading 32
4.10 High Level Abstraction 33
5 Experiment 34
5.1 Effectiveness 34
5.1.1 Optimization for Scene Parameters 34
5.1.2 Optimization for Surface Function 36
5.2 Efficiency 37
5.2.1 Speed 37
5.2.2 Memory Usage 39
5.3 Correctness of Curve Sampling 40
6 Conclusion 42
6.1 Future Work 43
6.2 Heuristic Optimization 43
6.2.1 Open Shading Language 44
6.2.2 Gradient Space Gathering 44
6.2.3 Gradient Space Denoising 44
6.2.4 Inverse Rendering 45
6.2.5 Deep Learning 45
Appendix A: Path Tracing 46
A.1 Bidirectional Reflectance Distribution Function 46
A.2 Infinite-Dimensional Integral of LTE 46
A.3 Ray-Quadric Intersection 47
A.4 Perspective Projection of Ellipsoid 48
Appendix B: Open Source Contribution 51
B.1 NVidia OptiX Prime for Python 51
B.2 Autodiff for Human 51
References 53
參考文獻 [1] Tzu-Mao Li. Differentiable monte carlo ray tracing through edge sampling. Siggraph Asia.
[2] Michael J. LOPER, Matthew M.; BLACK. Opendr: An approximate differentiable
renderer. pages 154–169, 2014.
[3] Paul Henderson and Vittorio Ferrari. Learning to generate and reconstruct 3d meshes with only 2d supervision. In British Machine Vision Conference (BMVC), 2018.
[4] Matthew Mirman. Mentisoculi. https://github.com/mmirman/MentisOculi, 2018.
[5] Yuanchun TAN, Ying; ZHU. Fireworks algorithm for optimization. pages 355–364, 2010.
[6] Volker BLANZ. A morphable model for the synthesis of 3d faces. Siggraph, pages
187–194, 1999.
[7] et al GKIOULEKAS, Ioannis. Inverse volume rendering with material dictionaries.
ACM Transactions on Graphics (TOG), 2013.
[8] James T Kajiya. he rendering equation. Siggraph, page 143–150, 1986.
[9] Donald P. Greenberg Cindy Goral, Kenneth E. Torrance and B. Modeling the interaction of light between diffuse surfaces.
[10] Henrik Wann JENSEN. Global illumination using photon maps. Rendering Techniques’ 96, pages 21–30.
[11] Eric VEACH. Robust monte carlo methods for light transport simulation. 1997.
[12] Marc ten Bosch. Python binding for nvidia optix prime. http://marctenbosch.com/photon/mbosch_intersection.pdf, 2004.
[13] Hsu YU-LUN. Python binding for nvidia optix prime. https://ppt.cc/fwJxIx, 2019.
論文全文使用權限
  • 同意授權校內瀏覽/列印電子全文服務,於2020-01-22起公開。
  • 同意授權校外瀏覽/列印電子全文服務,於2020-01-22起公開。


  • 如您有疑問,請聯絡圖書館
    聯絡電話:(06)2757575#65773
    聯絡E-mail:etds@email.ncku.edu.tw