Open long8v opened 1 year ago
paper
$w_A^{mn}$은 그냥 scaled dot attention이랑 비슷
$w_G^{mn}$는 feature를 뽑고(두개를 합쳐서 $\varepsilon_G$) sine/cosine으로 임베딩 시킨 뒤에 $W_g$ 곱해주고 ReLU 취해서 구해짐
뽑는 feature
최종적으로 $f^n_a$는 저렇게 뽑은 nm개의 object relation을 concat해서 나온다.
별거 아니고 그냥 {0, 1} 로 예측하는거. 근데 relation module이 있으니 중복을 잘 제거할 수 있을듯.
paper
TL;DR
Details
Object Relation Module
$w_A^{mn}$은 그냥 scaled dot attention이랑 비슷
$w_G^{mn}$는 feature를 뽑고(두개를 합쳐서 $\varepsilon_G$) sine/cosine으로 임베딩 시킨 뒤에 $W_g$ 곱해주고 ReLU 취해서 구해짐
뽑는 feature
최종적으로 $f^n_a$는 저렇게 뽑은 nm개의 object relation을 concat해서 나온다.
Relation for Instance Recognition
Relation for Duplicate Removal
별거 아니고 그냥 {0, 1} 로 예측하는거. 근데 relation module이 있으니 중복을 잘 제거할 수 있을듯.
Result