AdaFuse Adaptive Multiview Fusion for Accurate Human Pose Estimation in the Wild

enhance feature in occuluded viers by leveraging feature in visible views

• point-line corrspondence: point in one view correspond to a line in another view
  • aggragate multiple lines, find a optimal estimation from sampson distance
  • heatmap fusion
• adaptive weight for fusion
  • from appearance: some conv, mlp
  • from geometry: get sampson distance -> mlp
  • shared weight learning network