Landmark localization - 예시(facial landmark localization & Human pose estimation)
Predicting the coordinate of keypoints
1. Coordinate regression : Usually inaccurate and biased
2. heatmap classification : better performance but high computational cost
Hourglass network
stacked hourglass modules allow for repeated bottom-up and top-down inference that refines the output of the previous hourglass module
stack을 지날수록 더 refine 하게 추정
이전의 feature map의 결과를 convolution 한 것과 덧셈을 해서 새로운 layer로 활용함
UVmap(vector좌표 표기법) is a flattened representation of 3D geometry. Also, UV map is invariant to motion(i.e., cannonical coordinate)
DensePose R-CNN = Faster R-CNN + 3D surface regression branch
RetinaFace = FPN + Multi-task brances(classification, bounding box, 5 point regression, mesh regression)
동시에 다른 task를 학습하면서 공통된 대상의 다양한특징을 자연스럽게 학습하면서 서로 도와 성능향상을 도모할 수 있다.
요즘의 흐름 -> FPN + Target-task brances