Multiview geometry 1

Local Features

• 이미지에 중요한 부분
• Image representation
• Object appearance modeling
• Finding matches between multiple views

Model Fitting and RANSAC

• Finding vanishing points
• Image stitching
• 3D object recognition
• Fitting Technique: Least Square
• RANSAC
• Fitting Technique: Hough Transform

Geometric Transformations

• Understanding camera models
• 2D transformations
  • Translations
  • Euclidean transformation
  • Similarity transformation
  • Affine transformation
  • Projective transformation
• 3D Geometric Primitives
  • points, lines, planes
  • 3D Transformations
  • 3D to 2D Projections
• Camera Models
  • Camera
    • 사람의 눈에 대응되는 컴퓨터의 눈
  • Geometric Camera Models
    • Pinhole Model
      • Aperture shrinks to zero
      • Inverted image is observed in image plane

Camera Calibration

• Geometric camera calibra:on
• Camera Matrix Estimation
  • size and structure of the pattern are known
• Estimation of Intrinsic and Extrinsic Parameters

Two-View Geometry

• 2장의 이미지에서 대응 되는 기하학적 관계
• Homography
  • collinearity is preserved
  • Isotropic scaling transform
• Image Panorama Using Estimation

Epipolar Geometry

• epipolar
  • 두개의 이미지에서 대응되는 점들의 직선
• converging image planes
• parallel image planes
• forward moving camera
• Two-View Relationship in Epipolar Geometry
  • essential matrix
  • Fundamental Matrix
    • 8-point algorithm

Stereo Matching

• 두개의 센서로 부터 인지한 정보
• Parallel images
• Image Rectification
  • 같은 선상에 이미지를 정렬
  • Epipolar lines을 직선이 되도록 배치
• Basic Stereo Matching
  • epipolar line
  • depth vs disparity
  • 비워있으면 더 진하게 나옴
• Correspondence Search
• Correspondence Search by Correlation
• 반복적인 패턴
• small diparity error
• window size effect

Wide-Baseline Matching

• Large baseline
• KeyNet
  • detection 에 집중함
  • Multi-Scale Index Proposal (M-SIP) Loss.
• Self-Supervised Equivariant Learning for Oriented Keypoint Detection
  • Rotation-Equivariant Feature
  • Window-based keypoint loss
  • Dense orientation alignment loss

참조

---