学习文章:going deeper into action recognition : a survey。samitha herath 等。
传统方法:
wholistic:
local : sparse stip cuboid hog3d
dense trajectory
描述方法:(1)bow (zeroth order) (2)FV (zero,one,two,orderth information) (3)vlad(zero,one orderth information)
THUMOS challenge 2014 总结:
(1)dense trajectory features : 10 teams used DTF (9 improved DTF, 1 basic DTF)
(2)convolutional nerual netwroks: 4 teams
(3)beyond low-level features: body and facial features; audio; saliency features; shot boundary detection
(4)classification:10 teams used svm,1 extreme learning, no fully conneted cnn
(5)9 teams used fisher vector