상세정보
학술대회
Relabeling Method for Improving Vehicle Part Detection
- 발행일
- 202210
- 출처
- International Conference on Information and Communication Technology Convergence (ICTC) 2022, pp.806-808
- 협약과제
- 22ZH1200, 초실감 입체공간 미디어·콘텐츠 원천기술연구, 이태진
- 초록
- The lightweight deep learning detection method does not properly distinguish partial objects that are named differently depending on their location but has the same shape and appearance like a tire. To solve the problem, this paper introduces a method to relabel the detection result of vehicle parts. We define vehicles parts with super-class and sub-class according to the shape and location, and then detect super-classes. For relabeling super-class to sub-class, we classify vehicle viewing direction, and generate a set of sub-class combinations from super-classes. The spatial distributions among the detected partial objects are analyzed using the likelihood. Then, the labels of the detected parts are determined. We tested our method with self-collected and open dataset, and achieved a mAP value of 87.7%, which is about 11 % better than tiny YOLO v4.
- KSP 제안 키워드
- Detection Method, Part detection, Spatial distribution, deep learning(DL), open datasets, vehicle part