Post

3D Gaussian Splatting vs. NeRFs. What is the difference? πŸ€”

In the world of computer vision, 3D Gaussian Splatting and NeRFs are gaining traction.

But what sets them apart? Here’s a quick breakdown:

πŸ” 3D Space Representation:

  • NeRF: Creates a continuous 3D space by sampling points throughout the scene for each training image.
  • Gaussian Splatting: Relies on a sparse set of 3D points, often generated using Structure from Motion.

🎨 Points Description:

  • Gaussian Splatting: Uses complex 3D forms called Gaussians with varying shapes, sizes, transparency, and color, described with spherical harmonics.
  • NeRF: Assigns each point an RGBA color and a viewing direction, with appearance dependent on location and viewing angle.

βš™οΈ Optimization:

  • NeRF: Uses a neural network to learn a continuous function for color and opacity.
  • Gaussian Splatting: Directly optimizes the properties of each 3D ellipsoid without a neural network, resulting in a discrete set of ellipsoids.

In essence, NeRF offers a continuous, neural approach, while Gaussian Splatting provides a simpler, directly optimized discrete structure.

Which approach do you find more intriguing for applications in 3D graphics and AR? Share your thoughts in the comments! πŸ’¬

Information About 3D Gaussian Splatting

 3DGS Overview

Translate to Korean

컴퓨터 λΉ„μ „μ˜ μ„Έκ³„μ—μ„œλŠ” 3D Gaussian Splatting 및 NeRFκ°€ μ£Όλͺ©μ„ λ°›κ³  μžˆμŠ΅λ‹ˆλ‹€. κ·ΈλŸ¬λ‚˜ 무엇이 그듀을 μ°¨λ³„ν™”ν•©λ‹ˆκΉŒ? λ‹€μŒμ€ κ°„λ‹¨ν•œ λΆ„μ„μž…λ‹ˆλ‹€.

πŸ” 3D 곡간 ν‘œν˜„:

  • NeRF: 각 ν•™μŠ΅ 이미지에 λŒ€ν•΄ μž₯λ©΄ μ „μ²΄μ˜ 지점을 μƒ˜ν”Œλ§ν•˜μ—¬ 연속 3D 곡간을 λ§Œλ“­λ‹ˆλ‹€.
  • κ°€μš°μ‹œμ•ˆ μŠ€ν”Œλž˜νŒ…(Gaussian Splatting): μ’…μ’… λͺ¨μ…˜μ˜ ꡬ쑰(Structure from Motion)λ₯Ό μ‚¬μš©ν•˜μ—¬ μƒμ„±λ˜λŠ” ν¬μ†Œ 3D 포인트 μ„ΈνŠΈλ₯Ό μ‚¬μš©ν•©λ‹ˆλ‹€.

🎨 포인트 μ„€λͺ…:

  • κ°€μš°μ‹œμ•ˆ μŠ€ν”Œλž˜νŒ…: κ΅¬ν˜• 고쑰파둜 μ„€λͺ…λ˜λŠ” λ‹€μ–‘ν•œ λͺ¨μ–‘, 크기, 투λͺ…도 및 색상을 가진 κ°€μš°μ‹œμ•ˆμ΄λΌλŠ” λ³΅μž‘ν•œ 3D ν˜•νƒœλ₯Ό μ‚¬μš©ν•©λ‹ˆλ‹€.
  • NeRF: 각 ν¬μΈνŠΈμ— RGBA 색상과 보기 λ°©ν–₯을 ν• λ‹Ήν•˜λ©°, λͺ¨μ–‘은 μœ„μΉ˜ 및 μ‹œμ•Όκ°μ— 따라 λ‹¬λΌμ§‘λ‹ˆλ‹€.

βš™οΈ μ΅œμ ν™”:

  • NeRF: 신경망을 μ‚¬μš©ν•˜μ—¬ 색상 및 뢈투λͺ…도에 λŒ€ν•œ 연속 ν•¨μˆ˜λ₯Ό ν•™μŠ΅ν•©λ‹ˆλ‹€.
  • Gaussian Splatting: 신경망 없이 각 3D νƒ€μ›μ²΄μ˜ 속성을 직접 μ΅œμ ν™”ν•˜μ—¬ κ°œλ³„ 타원체 μ„ΈνŠΈλ₯Ό μƒμ„±ν•©λ‹ˆλ‹€.

본질적으둜 NeRFλŠ” 연속적인 μ‹ κ²½ μ ‘κ·Ό 방식을 μ œκ³΅ν•˜λŠ” 반면, Gaussian Splatting은 더 κ°„λ‹¨ν•˜κ³  직접 μ΅œμ ν™”λœ 이산 ꡬ쑰λ₯Ό μ œκ³΅ν•©λ‹ˆλ‹€.

3D κ·Έλž˜ν”½κ³Ό AR μ• ν”Œλ¦¬μΌ€μ΄μ…˜μ— μ–΄λ–€ μ ‘κ·Ό 방식이 더 ν₯λ―Έλ‘­λ‹€κ³  μƒκ°ν•˜μ‹­λ‹ˆκΉŒ? λŒ“κΈ€λ‘œ 생각을 κ³΅μœ ν•˜μ„Έμš”! πŸ’¬

This post is licensed under CC BY 4.0 by the author.