目录

目录

BAE-NET: Branched Autoencoder for Shape Co-Segmentation


<BAE-NET> BAE-NET: Branched autoencoder for shape co-segmentation

编者按

  • Learning Implicit Fields for Generative Shape Modeling (CVPR2019) 的续作,inside / outside indicator作为shape表征

Motivation

  • 把形状的 co-segmentation 看做表征学习问题
  • 可以无监督、弱监督、one-shot learning,只需要用几个exemplars,就可以在shape 分割任务上好过在分割shape上训练的SOTA
  • 无监督的 co-segmentation
    https://longtimenohack.com/posts/paper_reading/2019iccv_chen_bae/image-20201229094035378.png

overview

  • 就是在Learning Implicit Fields for Generative Shape Modeling 的基础上,从原来的单个inside / outside indicator变成 k 个inside / outside indicator (branched output, one neuron each) ,然后在最后max pooling 把几个neuron compose在一起。
  • https://longtimenohack.com/posts/paper_reading/2019iccv_chen_bae/image-20201229095002537.png
  • 让网络 “自动” 学出来一个个natural shape的neuron;没有强制保证
    https://longtimenohack.com/posts/paper_reading/2019iccv_chen_bae/image-20201229095340266.png