Multi-View correlation distillation for incremental object detection

In real applications, new object classes often emerge after the detection model has been trained on a prepared dataset with fixed classes. Fine-tuning old only data will lead to well-known phenomenon of catastrophic forgetting, which severely degrades performance modern detectors. Due storage burden, privacy and time consumption, sometimes it is impractical train from scratch all both this paper, we propose novel Multi-View Correlation Distillation (MVCD) based incremental method, explores intra-feature correlations in feature space detector. To better transfer knowledge learned maintain ability learn classes, select sample-specific discriminative features channel-wise, point-wise instance-wise views. Meanwhile, correlation distillation losses selective are designed regularize learning A metric named Stability-Plasticity-mAP (SPmAP) proposed evaluate as complementary mAP, integrates metrics for stability plasticity detection. The extensive experiments conducted VOC2007 COCO demonstrate that MVCD achieves trade-off between than state-of-the-art first-order distillation-based methods.