目录

目录

Unsupervised part representation by Flow Capsules


Unsupervised part representation by flow capsules

Motivation

  • capsule networks不能高效地学到low level的part descriptions
  • exploit motion as a powerful perceptual cue for part definition
    用运动作为一个部件定义的有力的感知线索

results

  • 从复杂背景中找出来原来的三角形、正方形、圆形等
    https://longtimenohack.com/posts/paper_reading/2021icml_sabour_unsupervised/image-20201216170936463.png
  • 对于运动的人学出来的部件https://longtimenohack.com/posts/paper_reading/2021icml_sabour_unsupervised/image-20201216171020883.png