Multi-View correlation distillation for incremental object detection

作者：

Highlights：

• A novel incremental object detection method is proposed to explore and transfer the multi-view correlations in the feature space of the object detector.

• The correlation distillation losses for the sample-specific selective features from three views (channel-wise, point-wise and instance-wise) are designed.

• A new metric called Stability-Plasticity-mAP (SPmAP) is proposed to measure the incremental object detector performance.

• The proposed method achieves competitive results compared with previous incremental object detection methods.

摘要

•A novel incremental object detection method is proposed to explore and transfer the multi-view correlations in the feature space of the object detector.•The correlation distillation losses for the sample-specific selective features from three views (channel-wise, point-wise and instance-wise) are designed.•A new metric called Stability-Plasticity-mAP (SPmAP) is proposed to measure the incremental object detector performance.•The proposed method achieves competitive results compared with previous incremental object detection methods.

论文关键词：Object detection,Incremental learning,Catastrophic forgetting,Knowledge distillation

论文评审过程：Received 27 October 2021, Revised 2 June 2022, Accepted 16 June 2022, Available online 18 June 2022, Version of Record 29 June 2022.

论文官网地址：https://doi.org/10.1016/j.patcog.2022.108863