핵심 개념
MultIOD is a class-incremental object detector based on CenterNet, focusing on rehearsal-free and anchor-free object detection.
초록
Class-Incremental Learning (CIL) is crucial in evolving environments.
Catastrophic forgetting challenges neural networks in retaining past knowledge.
MultIOD proposes a multihead feature pyramid and transfer learning to tackle forgetting.
Results show MultIOD outperforms state-of-the-art methods on Pascal VOC datasets.
Ablation studies highlight the impact of backbones, frozen layers, and NMS strategies.
MultIOD's efficiency lies in its reduced memory footprint and robustness against forgetting.
통계
대부분의 기존 CIOD 모델은 Faster-RCNN에 기반을 두고 있음.
MultIOD는 CenterNet 알고리즘을 기반으로 함.
EfficientNet-B3 모델은 파라미터 수를 줄이면서 성능을 향상시킴.
인용구
"We argue that more effort should be dedicated to anchor-free and rehearsal-free object detection."
"Results show that our method outperforms a range of state-of-the-art methods on two Pascal VOC datasets."