에어로졸 상황에서 물체의 6자유도 자세 추정을 위한 벤치마크 데이터셋 및 베이스라인 구축 (6-DoF Object Pose Estimation Under Aerosol Conditions: Benchmark Dataset and Baseline)

서지정보

· 발행기관 : 제어·로봇·시스템학회
· 수록지 정보 : 제어.로봇.시스템학회 논문지 / 30권 / 6호 / 614 ~ 620페이지
· 저자명 : 양희진, 이승현, 김태주, 최유경

초록

Research on 6-degrees of freedom (6-DoF) pose estimation has been conducted in several fields until recently. As deep learning advances, the corresponding 6-DoF pose estimations will be trained on a variety of data to make predictions after training. However, research on using datasets to perform 6-DoF pose estimation in environments such as those involving aerosols, which can be encountered in disaster areas, is presently lacking. In this paper, we propose a novel RGB-D benchmark dataset composed of paired images under both normal and aerosol conditions, also offering a baseline. We specifically address the overall pipeline, i.e., 3D model construction, sensor setup, implementation of data collection methods, the 6-DoF pose annotation process applicable to objects, and the validation of these annotations, to build the dataset proposed in this study. Using our proposed dataset, we apply previously studied 6-DoF pose estimation methodologies to benchmark and experiment on normal and aerosol situations. Our experiments reveal a substantial performance degradation in aerosol situations, and further experiments to investigate the aerosol environment reveal that performance can be improved. The aerosol conditions demonstrate the difficulty of performing 6-DoF pose estimation and the need for future research.

영어초록

Research on 6-degrees of freedom (6-DoF) pose estimation has been conducted in several fields until recently. As deep learning advances, the corresponding 6-DoF pose estimations will be trained on a variety of data to make predictions after training. However, research on using datasets to perform 6-DoF pose estimation in environments such as those involving aerosols, which can be encountered in disaster areas, is presently lacking. In this paper, we propose a novel RGB-D benchmark dataset composed of paired images under both normal and aerosol conditions, also offering a baseline. We specifically address the overall pipeline, i.e., 3D model construction, sensor setup, implementation of data collection methods, the 6-DoF pose annotation process applicable to objects, and the validation of these annotations, to build the dataset proposed in this study. Using our proposed dataset, we apply previously studied 6-DoF pose estimation methodologies to benchmark and experiment on normal and aerosol situations. Our experiments reveal a substantial performance degradation in aerosol situations, and further experiments to investigate the aerosol environment reveal that performance can be improved. The aerosol conditions demonstrate the difficulty of performing 6-DoF pose estimation and the need for future research.

참고자료

· 없음