(NeurIPS 2015) Spatial Transformer Networks

* 논문 링크 : https://arxiv.org/pdf/1506.02025v3.pdf



아직은 논문을 읽기 전 상태로, 아래 링크의 블로그를 번역 정리한 내용을 작성하였음

논문은 추후 읽고 정리할 예정

https://towardsdatascience.com/review-stn-spatial-transformer-network-image-classification-d3cbd98a70aa



Spatial Transformer Networks는 Google DeepMind에서 만든 Network임

적절한 영역을 자르고 스케일 정규화하는 데 도움이 되어 분류 작업을 단순화하고 더 나은 성능을 보여줌



본 논문은 Spatial Transform을 신경망으로 처리함

학습 기반 공간 변환에서 입력 또는 feature map을 조건으로 변환이 적용됨



본 논문과 관련 논문으로 Deformable Convolutional Networks이 있음



[Spatial Transformation Matrix]

1. Affine Transformation

 - 행렬 값에 따라 다양한 효과를 사용해 (X1, Y1)을 (X2, Y2)로 변환 가능

 - Translate, Scale, Rotate, Sheer



2. Projective Transformation

 - 사영 변환도 학습 가능



3. Thin Plate Spline (TPS) Transformation

 

[Spatial Transformer Network]

 - Localization Net, Grid Generator, Sampler로 구성됨



1. Lacalization Net

- affine 변환으로 학습 

- 출력은 $\theta$



2. Grid Generator

- 점 집합



3. Sampler

- 변환된 그리드를 샘플링해야하는 경우 샘플링 문제가 발생함 

- 이를 보완해주기 위함