yiskw713 / paper_summary

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

ActionFlowNet: Learning Motion Representation for Action Recognition #151

Open yiskw713 opened 4 years ago

yiskw713 commented 4 years ago

INFO

author

Joe Yue-Hei Ng et al.

affiliation

conference or year

WACV 2018

link

arXiv

概要

Screen Shot 2020-05-25 at 12 28 27

Optical Flowの生成と行動認識を同時に行うモデルである ActionFlowNetを提案. 上図と下図の二つのモデルを提案しているが,上図の方が性能が良い.

Screen Shot 2020-05-25 at 12 28 56

提案手法

検証

精度的にはI3Dや,ImageNet-pretrained ResNetの方が高い. Screen Shot 2020-05-25 at 12 29 47

一方でCAMを比較すると,提案手法の方がより動きの特徴に着目していることがわかる. Screen Shot 2020-05-25 at 12 29 15

date