long8v / PTIR

Paper Today I Read
19 stars 0 forks source link

[88] Relation Networks for Object Detection #97

Open long8v opened 1 year ago

long8v commented 1 year ago
image

paper

TL;DR

Details

image

Object Relation Module

image image image

$w_A^{mn}$은 그냥 scaled dot attention이랑 비슷

image

$w_G^{mn}$는 feature를 뽑고(두개를 합쳐서 $\varepsilon_G$) sine/cosine으로 임베딩 시킨 뒤에 $W_g$ 곱해주고 ReLU 취해서 구해짐

image

뽑는 feature

image

최종적으로 $f^n_a$는 저렇게 뽑은 nm개의 object relation을 concat해서 나온다.

Relation for Instance Recognition

image

Relation for Duplicate Removal

image

별거 아니고 그냥 {0, 1} 로 예측하는거. 근데 relation module이 있으니 중복을 잘 제거할 수 있을듯.

Result

image image image