핵심 개념
Proposing a novel sparsely-annotated framework for 3D object detection to reduce annotation burden while maintaining performance.
초록
The article discusses the challenges of densely-annotated 3D object detection datasets and proposes a sparse annotation strategy to reduce annotation costs. It introduces the SS3D++ method that progressively generates confident fully-annotated scenes based on sparse annotations. The method achieves competitive results with less annotation costs compared to weakly-supervised methods and on-par performance with fully-supervised methods. The article also highlights the importance of additional unlabeled training scenes in boosting performance.
통계
현재 최첨단 3D 객체 감지 방법은 대규모 3D 경계 상자 주석을 필요로 함.
SS3D++ 방법은 신뢰할 수 있는 완전 주석된 장면 생성을 통해 경쟁력 있는 결과 달성.
KITTI 데이터셋에서 5배 적은 주석 비용으로 SOTA 완전 지도 방법과 유사한 성능 달성.
Waymo 데이터셋에서 15배 적은 주석 비용으로 90%의 성능 달성.
인용구
"To reduce the cumbersome data annotation process, we propose a novel sparsely-annotated framework."
"Our proposed method produces competitive results when compared with SOTA weakly-supervised methods."