Notice
Recent Posts
Recent Comments
Link
«   2024/09   »
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
Tags
more
Archives
Today
Total
관리 메뉴

what i learned

[논문리뷰] CenterNet(Objects as Point) 본문

Paper Review

[논문리뷰] CenterNet(Objects as Point)

햄식이111 2023. 9. 21. 23:56
 

GitHub - xingyizhou/CenterNet: Object detection, 3D detection, and pose estimation using center point detection:

Object detection, 3D detection, and pose estimation using center point detection: - GitHub - xingyizhou/CenterNet: Object detection, 3D detection, and pose estimation using center point detection:

github.com

 

CenterNet이란?

CenterNet은 객체를 검출하는 모델로,

(1)Bbox를 통해 객체를 검출하는 방법이 아닌 (2)Key Point를 통해 객체를 검출하는 방법이다.

 

CenterNet의 차별성

이 논문에서는 Bbox를 통한 객체를 검출하는 방법이 비효율적이라 Key Point를 사용하여 검출한다고 말한다.

  1. Bbox overlap이 아니라 위치에 기반한 anchor를 할당
    (배경인지 아닌지 분류하는 임계값이 존재하지 않아도 된다.)
  2. 객체 하나당 positive anchor만 있어서 NMS 불필요
    (NMS가 아닌 key point heatmap에서 local peak를 추출한다.)
  3. 큰 output resolution을 사용하여 여러개의 anchor 불필요
    (기존 output stride를 16으로 사용하였지만 stride를 4로 줄여 resolution을 높였다.)

 

Key Point 검출

CenterNet에서 KeyPoint를 검출하는 방법은 위와 같다.

바운딩 박스의 상단 왼쪽과 하단 오른쪽의 x, y좌표 x1,x2,y1,y2를 가지고 바운딩 박스 내 중점을 구한다.

그 후 2개의 좌표를 통해 bbox의 사이즈도 유추할 수 있다.

 

사실 key point를 가지고 객체를 검출하는 모델은 이번이 처음이 아니다.

CornerNet과 ExtremeNet 모델도 key point를 사용하여 객체를 검출하지만,

CornerNet은 바운딩박스의 왼쪽 상단과 오른쪽 하단 2개의 key point를 사용하여 객체를 검출하고

ExtremeNet은 5개 이상의 keypoint를 검출하기 때문에

단 1개의 key point만을 사용하여 객체를 검출하는 CenterNet은 다른 key point 모델들에 비해 연산량이 적어 속도가 빠르다는 장점이 있다.