Taskonomy: Disentangling Task Transfer Learning

cvpr2018最佳paper，其目标在于利用视觉任务之间的关联性减少总体数据使用量。

• paper
• 作者解读
• github

感兴趣的问题：

• 哪26个任务

  • surface normals
  • Eucl.distance
  • object class.
  • scene class.
  • jigsaw puzzle
  • colorization
  • 2D segm.
  • 2.5D segm.
  • semantic segm.
  • valishing points
  • 2D edges
  • 3D edges
  • 2D keypoints
  • 3D keypoints
  • 3D curvature
  • image reshading
  • in-painting
  • denoising
  • autoencoding
  • cam. pos(non fixed)
  • cam. pos(fixed)
  • triplet cam pose
  • room layout
  • point matching
  • random projection

  • Egomotion(帧间运动)

    
    
    tasks.png

• 与semantic segmentation相关的任务有哪些

  • object class.

  • scene class.

    
    
    task tree.png

• 网络结构：类ResNet50