3D Gaussian Splatting for
Real-Time Radiance Field Rendering
(SIGGRAPH 2023)
이미지처리팀
김병현(발표자), 김준철, 최승준, 허정원
Sep 17th, 2023
Gaussian Splatting
 High-quality and real-Time radiance field rendering method (≥ 100
fps at 1080p) based on 3D Gaussians
2
Gaussian Splatting
 High-quality and real-Time radiance field rendering method (≥ 100
fps at 1080p) based on 3D Gaussians
3
Gaussian Splatting
 High-quality and real-Time radiance field rendering method (≥ 100
fps at 1080p) based on 3D Gaussians
4
RWs-1: Traditional Scene Reconstruction
 Lumigraph
5
Gortler, S. J., Grzeszczuk, R., Szeliski, R., & Cohen, M. F. (2023). The lumigraph.
In Seminal Graphics Papers: Pushing the Boundaries, Volume 2 (pp. 453-464).
The plenoptic function is a five dimensional
quantity describing the flow of light at
every 3D spatial position (x, y, z) for every 2D
direction (θ,Φ).
RWs-1: Traditional Scene Reconstruction
 Structure from Motion
6
Snavely, N., Seitz, S. M., & Szeliski, R. (2006). Photo tourism: exploring photo
collections in 3D. In ACM siggraph 2006 papers (pp. 835-846).
RWs-1: Traditional Scene Reconstruction
 Structure from Motion
7
Snavely, N., Seitz, S. M., & Szeliski, R. (2006). Photo tourism: exploring photo
collections in 3D. In ACM siggraph 2006 papers (pp. 835-846).
RWs-1: Traditional Scene Reconstruction
 Multi-View Stereo for Community Photo Collections
8
Goesele, M., Snavely, N., Curless, B., Hoppe, H., & Seitz, S. M. (2007, October). Multi-view stereo for community
photo collections. In 2007 IEEE 11th International Conference on Computer Vision (pp. 1-8). IEEE.
RWs-2: Neural Rendering & Radiance Fields
 View Synthesis by Appearance Flow
9
Zhou, T., Tulsiani, S., Sun, W., Malik, J., & Efros, A. A. (2016). View synthesis by appearance flow. In Computer Vision–ECCV 2016: 14th European
Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14 (pp. 286-301). Springer International Publishing.
RWs-2: Neural Rendering & Radiance Fields
 NeRF: Representing Scenes as Neural Radiance Fields for View
Synthesis
10
Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., & Ng, R. (2021). Nerf: Representing scenes
as neural radiance fields for view synthesis. Communications of the ACM, 65(1), 99-106.
RWs-2: Neural Rendering & Radiance Fields
 Mip-Nerf
11
Barron, J. T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., & Srinivasan, P. P. (2021). Mip-nerf: A multiscale representation
for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 5855-5864).
RWs-2: Neural Rendering & Radiance Fields
 Instant Neural Graphics Primitives with a Multiresolution Hash
Encoding
12
Müller, T., Evans, A., Schied, C., & Keller, A. (2022). Instant neural graphics primitives with a
multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4), 1-15.
RWs-3: Point-Based Rendering
13
Sainz, M., & Pajarola, R. (2004). Point-based rendering
techniques. Computers & Graphics, 28(6), 869-879.
RWs-3: Point-Based Rendering
14
Yifan, W., Serena, F., Wu, S., Öztireli, C., & Sorkine-Hornung, O. (2019). Differentiable surface
splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6), 1-14.
Q & A
Q & A
15
Method
16
3D Gaussians
 A Gaussian distribution on 3D space with centered at point (mean) 𝜇
and a full 3D covariance matrix Σ
17
https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/
3D Gaussian Visualization:
Left side of each image is the final rendering result, and the right side is 3D Gaussians
3D Gaussians
 A Gaussian distribution on 3D space with centered at point (mean) 𝜇
and a full 3D covariance matrix Σ
18
https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/
Adaptive Density Control
19
Fast Differentiable Rasterization
 Rasterization
20
https://www.scratchapixel.com/lessons/3d-basic-rendering/rasterization-
practical-implementation/overview-rasterization-algorithm.html
Concept of rasterization
Fast Differentiable Rasterization
21
Fast Differentiable Rasterization
22
Fast Differentiable Rasterization
23
Optimization
24
Q & A
Q & A
25
Datasets
26
Tank & Temple (150-300 Images)
(https://www.tanksandtemples.org/)
MIP-Nerf 360 (100-330 Images)
(https://jonbarron.info/mipnerf360/)
Deep Blending
(http://visual.cs.ucl.ac.uk/pubs/deepblending/)
Experiments
 Quantitative Comparison
27
Quantitative evaluation compared to previous work
Experiments
 Quantitative Comparison
28
Quantitative evaluation compared to previous work
Experiments
 Quantitative Comparison
29
Quantitative evaluation compared to previous work
Experiments
 Qualitative Comparison
30
Ablation Study
 Optimization Iteration
31
Ablation Study
 Initialization with SfM points
32
Ablation Study
 Initialization with SfM points
33
Shrunken Gaussians
34
Limitations
35
Q & A
Q & A
36
Thank you for your attention!
Thank you for your attention!
37

3D Gaussian Splatting

  • 1.
    3D Gaussian Splattingfor Real-Time Radiance Field Rendering (SIGGRAPH 2023) 이미지처리팀 김병현(발표자), 김준철, 최승준, 허정원 Sep 17th, 2023
  • 2.
    Gaussian Splatting  High-qualityand real-Time radiance field rendering method (≥ 100 fps at 1080p) based on 3D Gaussians 2
  • 3.
    Gaussian Splatting  High-qualityand real-Time radiance field rendering method (≥ 100 fps at 1080p) based on 3D Gaussians 3
  • 4.
    Gaussian Splatting  High-qualityand real-Time radiance field rendering method (≥ 100 fps at 1080p) based on 3D Gaussians 4
  • 5.
    RWs-1: Traditional SceneReconstruction  Lumigraph 5 Gortler, S. J., Grzeszczuk, R., Szeliski, R., & Cohen, M. F. (2023). The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2 (pp. 453-464). The plenoptic function is a five dimensional quantity describing the flow of light at every 3D spatial position (x, y, z) for every 2D direction (θ,Φ).
  • 6.
    RWs-1: Traditional SceneReconstruction  Structure from Motion 6 Snavely, N., Seitz, S. M., & Szeliski, R. (2006). Photo tourism: exploring photo collections in 3D. In ACM siggraph 2006 papers (pp. 835-846).
  • 7.
    RWs-1: Traditional SceneReconstruction  Structure from Motion 7 Snavely, N., Seitz, S. M., & Szeliski, R. (2006). Photo tourism: exploring photo collections in 3D. In ACM siggraph 2006 papers (pp. 835-846).
  • 8.
    RWs-1: Traditional SceneReconstruction  Multi-View Stereo for Community Photo Collections 8 Goesele, M., Snavely, N., Curless, B., Hoppe, H., & Seitz, S. M. (2007, October). Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision (pp. 1-8). IEEE.
  • 9.
    RWs-2: Neural Rendering& Radiance Fields  View Synthesis by Appearance Flow 9 Zhou, T., Tulsiani, S., Sun, W., Malik, J., & Efros, A. A. (2016). View synthesis by appearance flow. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14 (pp. 286-301). Springer International Publishing.
  • 10.
    RWs-2: Neural Rendering& Radiance Fields  NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis 10 Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., & Ng, R. (2021). Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1), 99-106.
  • 11.
    RWs-2: Neural Rendering& Radiance Fields  Mip-Nerf 11 Barron, J. T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., & Srinivasan, P. P. (2021). Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 5855-5864).
  • 12.
    RWs-2: Neural Rendering& Radiance Fields  Instant Neural Graphics Primitives with a Multiresolution Hash Encoding 12 Müller, T., Evans, A., Schied, C., & Keller, A. (2022). Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4), 1-15.
  • 13.
    RWs-3: Point-Based Rendering 13 Sainz,M., & Pajarola, R. (2004). Point-based rendering techniques. Computers & Graphics, 28(6), 869-879.
  • 14.
    RWs-3: Point-Based Rendering 14 Yifan,W., Serena, F., Wu, S., Öztireli, C., & Sorkine-Hornung, O. (2019). Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6), 1-14.
  • 15.
    Q & A Q& A 15
  • 16.
  • 17.
    3D Gaussians  AGaussian distribution on 3D space with centered at point (mean) 𝜇 and a full 3D covariance matrix Σ 17 https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ 3D Gaussian Visualization: Left side of each image is the final rendering result, and the right side is 3D Gaussians
  • 18.
    3D Gaussians  AGaussian distribution on 3D space with centered at point (mean) 𝜇 and a full 3D covariance matrix Σ 18 https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/
  • 19.
  • 20.
    Fast Differentiable Rasterization Rasterization 20 https://www.scratchapixel.com/lessons/3d-basic-rendering/rasterization- practical-implementation/overview-rasterization-algorithm.html Concept of rasterization
  • 21.
  • 22.
  • 23.
  • 24.
  • 25.
    Q & A Q& A 25
  • 26.
    Datasets 26 Tank & Temple(150-300 Images) (https://www.tanksandtemples.org/) MIP-Nerf 360 (100-330 Images) (https://jonbarron.info/mipnerf360/) Deep Blending (http://visual.cs.ucl.ac.uk/pubs/deepblending/)
  • 27.
    Experiments  Quantitative Comparison 27 Quantitativeevaluation compared to previous work
  • 28.
    Experiments  Quantitative Comparison 28 Quantitativeevaluation compared to previous work
  • 29.
    Experiments  Quantitative Comparison 29 Quantitativeevaluation compared to previous work
  • 30.
  • 31.
  • 32.
  • 33.
  • 34.
  • 35.
  • 36.
    Q & A Q& A 36
  • 37.
    Thank you foryour attention! Thank you for your attention! 37