Notice
Recent Posts
Recent Comments
Link
일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | 5 | ||
6 | 7 | 8 | 9 | 10 | 11 | 12 |
13 | 14 | 15 | 16 | 17 | 18 | 19 |
20 | 21 | 22 | 23 | 24 | 25 | 26 |
27 | 28 | 29 | 30 |
Tags
- optimizationalgorithms
- 3dscanning #digitaltwintechnology #3dcontentcreation
- turtlebot3 #터틀봇
- anisotropicgaussianlearning #gaussiansplatting #nonisotropicrenderinㅎ
- machinelearning3d #deeplearninggraphics #airendering
- advancedrenderingtechniques #neuralscenerepresentation
- computergraphics #3dmodeling #virtualreality #augmentedreality #gamedevelopment
- pointbasedrendering #computergraphics #3dmodeling #volumerendering
- 3dpointcloud #differentiablerendering #3dscenerepresentation #neuralrendering
- anisotropickernels #ellipsoidalsplatting #orientedsplats #gradientbasedlearning
- siggraphtechniques #aigraphics #3dmodelingalgorithms
- gaussiansplatting #3dgaussiancovariance #nerf #3dreconstruction
- gaussiansplatting #3dgaussiansplatting #pointbasedrendering
- highfidelityreconstruction #sceneunderstanding #computationalgraphics
- gaussianprojection #covariancematrix3d #anisotropicgaussians #ellipsoidalsplatting
- splatrendering #3dpointcloud #differentiablerendering
- 3dcontentcreation
- nextgengraphics #futureof3drendering #innovativerenderingtechniques
- computervisionresearch #3dreconstructiontechniques #graphicsresearch
- differentiablerendering #machinelearning3d #deeplearninggraphics
- 3dscenerepresentation #covariancematrixoptimization #adaptivegaussians
- realtimerendering #highquality3d #fastrendering
- nerf (neural radiance fields) #3dreconstruction
- gpuacceleration #aigraphics #virtualreality #augmentedreality #gamedevelopment
- geometrylearning #shapeoptimization #gpuacceleration #realtimerendering
- 3drenderingtools #pointcloudsoftware #3dvisualizationsoftware
- 3dsceneunderstanding #pointcloudrendering #neuralscenerepresentation
- realtimerendering #machinelearning3d #deeplearninggraphics #computervision
- siggraph #3dsceneunderstanding #highquality3drendering #fastrendering
- nerf (neural radiance fields) #3dreconstruction #pointcloudrendering #volumerendering
Archives
- Today
- Total
Wiredwisdom
3D Gaussian's Split conditions 본문
위는 가우시안의 크기와 Opacity를 관장하는 파라미터들에 대한
Back-propagation 학습의 관계를 설명했다.
다음은 각 가우시안들이 어떤 조건에서 분기 되는지를 알아보자.
각 가우시안의 Opacity는 가우시안의 최대 밀도 부분에서의 값이고
밀도의 감소에 따라 Opacity의 값도 감소한다.
다음은 가우시안의 분기 조건이다.
조건은 두개다.
하나는 지역적으로 밀도가 기준치 이상으로 높은 경우다. 가우시안의 밀도를 의미하지 않는다.
또 다른 하나는 가우시안의 크기가 기준치 이상으로 커질 경우이다.
이 두가지 경우에 가우시안은 분기가 이루어지며, 분기된 가우시안의 크기와 위치는
기존 가우시안의 주축을 기점으로 분포되게 된다.
크기는 1/2, 2/3 정도로 줄어들며 이는 고정할 수도 있지만, 적응적으로 변하도록 계획할 수도 있다.
<색상은 상속되지 않는다>
장점:
- 계산 효율성:
- 축 정렬 구조는 연산이 간단하고 빠릅니다.
- 특히 GPU 아키텍처에 최적화되어 있어 병렬 처리에 유리합니다.
- 메모리 효율성:
- 축 정렬 구조는 저장과 접근이 단순해 메모리 사용을 최적화할 수 있습니다.
- 구현의 용이성:
- 알고리즘 구현이 상대적으로 간단해집니다.
- 디버깅과 최적화가 더 쉬워집니다.
- 빠른 수렴:
- 단순한 구조로 인해 학습 초기 단계에서 빠른 수렴이 가능할 수 있습니다.
- 계층적 구조와의 호환성:
- 옥트리(Octree)와 같은 계층적 데이터 구조와 자연스럽게 통합됩니다.
- 샘플링 효율성:
- 축 정렬 구조는 레이 캐스팅이나 샘플링 과정을 단순화할 수 있습니다.
단점:
- 표현력의 제한:
- 비축정렬 특성을 가진 표면을 표현하는 데 더 많은 가우시안이 필요할 수 있습니다.
- 오버피팅 가능성:
- 복잡한 구조를 표현하기 위해 과도하게 많은 가우시안을 사용하면 오버피팅의 위험이 있습니다.
- 메모리 트레이드오프:
- 복잡한 구조를 표현하기 위해 더 많은 가우시안을 사용하면 결과적으로 메모리 사용량이 증가할 수 있습니다.
결론적으로, 축 방향 성장 및 분할 방식은 계산 효율성과 구현 용이성 측면에서 큰 이점을 제공합니다. 이는 특히 실시간 렌더링이나 대규모 장면 표현에 중요합니다. 학습 이터레이션 수가 증가할 수 있지만, 각 이터레이션의 계산 비용이 낮아 전체적인 학습 시간은 크게 영향받지 않을 수 있습니다.
'Vision > Gaussian Splatting' 카테고리의 다른 글
Gaussian-splatting test - basic setup (0) | 2025.01.08 |
---|---|
가우시안 스플래팅의 비등방성 학습 해결 방안 (0) | 2024.07.10 |
3D Gaussian-Covariance (0) | 2024.07.03 |
FAST DIFFERENTIABLE RASTERIZER FOR GAUSSIANS (0) | 2024.06.30 |
Convert 3D Gaussian to 2D splat Method (0) | 2024.06.22 |