HOTA: High Order Tracking Accuracy

1. 등장 배경

• MOTA와 IDF1의 아래와 같은 문제점을 보완하기 위함
  • detection과 association을 지나치게 강조
  • predicted box와 GT box 사이의 detection similarity threshold를 0.5로 고정시켜서 값을 측정

 

2. 개념

• detection score와 association score의 조합으로, accurate detection과 association의 균형을 맞춘 것
• detection similarity threshold를 0.05 부터 0.95까지 0.05 intervals로 값을 바꿔가면서 측정
  • localization accuracy를 더 잘 고려

 

3. 계산 방법

• $\text {HOTA}_{\alpha } = \sqrt{\frac{\sum _{c \in \{\text {TP}\}} \mathcal {A}(c) }{|\text {TP}| + |\text {FN}| + |\text {FP}|}}$
  • $\mathcal {A}(c) = \frac{|\text {TPA}(c)|}{|\text {TPA}(c)| + |\text {FNA}(c)| + |\text {FPA}(c)|}$    ($**\mathcal {A}(c)$ is Association)
    • $TPA$ : $c$와 동일한 gtID와 prID를 모두 가진 $TP$의 집합
    • $FNA$ : $c$와 동일한 gtID를 가진 집합이지만, $c$와 다른 prID를 할당하였거나 누락한 경우
    • $FPA$ : $c$와 동일한 prID를 가진 집합이지만, $c$와 다른 gtID를 할당하였거나 object가 아닌 경우

 

참고 링크

https://arxiv.org/abs/2009.07736