yiskw713 / paper_summary

Write summaries of papers I've read in Japanese
10 stars 1 forks source link

The Lovasz-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks #84

Open yiskw713 opened 4 years ago

yiskw713 commented 4 years ago

INFO

author

Maxim Berman, Amal Rannen Triki, Matthew B. Blaschko

affiliation

Dept. ESAT, Center for Processing Speech and Images

conference or year

CVPR2018

link

pdf 実装

概要

IOUを最大化するロス関数を提案.Jaccard Loss は離散的であり最適化がうまく行かないことが多い.この問題を解決するために,Jaccard Loss に Lovasz-extention を行った. 解説記事が非常にわかりやすいです.

date

Oct. 3rd, 2019