Notice
Recent Posts
Recent Comments
Link
«   2024/10   »
1 2 3 4 5
6 7 8 9 10 11 12
13 14 15 16 17 18 19
20 21 22 23 24 25 26
27 28 29 30 31
Archives
Today
Total
관리 메뉴

Deep CV

ATSS : Adaptive Training Sample Selection 알고리즘 간단 정리 본문

딥러닝/Object Detection & Segmentation

ATSS : Adaptive Training Sample Selection 알고리즘 간단 정리

Present_Kim 2021. 12. 4. 19:49

Object Detection에서는 BBOX의 positive sample과 Negative sample을 정의하는 알고리즘이 중요합니다.

ATSS는 Hyperparameter가 필요 없다.

1. g(ground-truth box)의 positive sample 후보를 정한다.

  • 각 pyramid level 마다 g와 가장 가까운 k개의 anchor box를 구한다.
  • g의 center 그리고 anchor box의 center와의 L2 distance 이용
  • 결국 하나의 ground-truth box g k×L개의 positive sample 후보(Cg)가 생긴다.

2. ground truth g와 후보들(Cg)간의 IoU를 계산한다=Dg. 그 후 mg=mean(Dg)과 Vg=standard deviation(Dg)을 구한다. 

 

3. ground truth g와의 IoU가 특정 threshold(tg = mg+vg) 값 보다 큰 후보를 최종 positive sample(P)로 선택한다.

- (line 11 ~ 15) 

  • positive sample의 center가 ground truth g안에 있는 경우에만 선택
  • threshold tg = mg+vg
  • 하나의 anchor box가 여러 ground-truth box의 positive sample이 된다면, 가장 높은 IoU를 가진 쪽으로 선택된다.

4. 전체 anchor box에서 positive sample로 선택 받지 못한 anchor들은 negative sample이 된다. - (line 17)