Dynamic Augmentation Strategy Selection for Incremental Object Detection

  • Oh, Yujeong; 
  • Kim, Taehoon; 
  • Lee, Mingyu; 
  • Yun, Kimin; 
  • Choi, Jongwon
Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Incremental Object Detection (IOD) suffers from the challenges of balancing the capacity to learn new information and the ability to retain prior knowledge, further facing challenges such as missing annotations. Although an image can contain multiple classes of objects, annotations are provided only for the subset of classes given at each step, making IOD a particularly challenging problem in the realm of incremental learning. Recent studies suggest the efficacy of augmented bounding box replay in alleviating missing annotations, but these fail to offer flexible augmentation strategy selection. Our preliminary experiments reveal that the effective augmentation strategies differ by class, highlighting the limitations of existing fixed augmentation strategy selection methods. To address this issue, we propose a dynamic module for determining effective augmentation strategies. Moreover, we introduce an “Old label to Background” (OB) loss that effectively utilizes old class labels in augmented data that the conventional unbiased loss overlooks. We perform extensive experimental validation using the Pascal VOC dataset, showing the effectiveness of our selection scheme.

키워드

Incremental learning; Incremental object detection; Object detection
제목
Dynamic Augmentation Strategy Selection for Incremental Object Detection
저자
Oh, Yujeong; Kim, Taehoon; Lee, Mingyu; Yun, Kimin; Choi, Jongwon
DOI
10.1007/978-3-032-31930-2_44
발행일
2027
유형
Proceedings Paper
저널명
Lecture Notes in Computer Science
권
16825
페이지
654 ~ 668