microexpression recognition.
This repo is to realize short time mirco expression recognition in realtime. Base model is LSTM and dataset for pretrain is CK+.
Pretrain is 10 fold train.
hn of lstm can be regarded as video level feature.
self attention weight is α = σ(linear(lms)).
relation attention weight is β = σ(linear([lms:hn])).
so, a sample aggregated feature is Σαβ[lms:hn]/Σαβ.