Multiview geometry 2

Wide-Baseline Matching

Patch-based model

• L2-Net
• HardNet
• SOSNet
• HyNet

Dense Imaging Model

• GIFT
• GLU-Net

Joint Detection and Description

• LIFT
• LF-Net
• SuperPoint
  • Self-supervised for local features
  • Loss funcVons
    • Detector loss + Descriptor loss
• D2-NET
  • Feature descriptors: 이전에 DELF 나 CNNGeometric 처럼 Dense CNN Feature를 local descriptors로 본다. 이 descripctor vector는 Euclidean distance를 계산할 준비가 된 상태이다. 실제로는 채널 방향의 L2 norm을 한다.
  • feature detectors: raw CNN feature의 각 채널의 Post-processing을 통해 구한다
• R2D2

Local Regional Information Estimation

• OriNet
  • Learning to assign the local orientation values in the image matching pipeline
• AffNet
  • Learning local affine shape estimator
• Self-supervised learning of image scale and orientation estimation : SelfScaOri

Matching Models

• Learning to find good correspondences
• SuperGlue
  • context aggregation + matching + filtering

End-to-End Models

• LoFTR
  • Based on transformer blocks
• COTR
  • Correspondence Transformer
• DKM

참조

---