Incremental Object Detection Inspired by Memory Mechanisms in Brain

Incremental learning is key to bridging the enormous gap between artificial intelligence and human intelligence,mea-ning that agents can learn several tasks sequentially from a continuous stream of correlated data without forgetting,just as humans do.Object detection is one of the core tasks in the field of computer vision and the cornerstone of computer images understanding.Therefore,the incremental object detection has important research and practical significance.Although incremental learning has achieved good results in image classification,the research on incremental learning based on object detection is still in its infancy.This is because object detection is more complex than image classification,which needs to solve both classification and bounding box regression problems.Many researchers have made great efforts to solve this problem,but most of the work only focuses on how to retain previous learning,ignoring fast adaptability to new tasks,which is a critical requirement for incremental learning.Based on the memory mechanism of the brain,humans can constantly extract knowledge during learning,so as to learn new tasks better and faster without forgetting.Inspired by this,an incremental meta-learning method that integrates the codec memory replay mechanism is proposed.This method encodes,stores,decodes and replays the feature vectors of learned samples,so as to approximate the dynamic learning environment as a local stationary environment and avoid catastrophic forgetting.Besides,a double-loop online meta-learning strategy is designed,which can help model to extract common structures of tasks and improve generalization performance on new tasks encountered during learning.The model is respectively updated by SGD with multiple batches of old and new mixed data in the inner loop,and is meta-updated in the outer loop.We evaluate the proposed approach on three incremental object detection settings defined on PASCAL VOC and MS COCO datasets,where the proposed algorithm performs favorably well against state-of-the-art methods.It proves that it can help the model to resist forgetting better and have better generalization performance on new tasks.The proposed algorithm is gradient-based and model-agnostic,so it has strongadaptability and can be applied on more complex detection frameworks.