Incremental Learning for Object Detection Based on Self-Supervised Learning

Incremental learning, as an approach well-suited for handling applications characterized by evolving data streams, aligns effectively with the objectives of object detection. However, existing methods exhibit suboptimal performance in addressing issues related to catastrophic forgetting and few-shot learning of new classes. Through the enhancement of existing models, we propose a novel method that integrates self-supervised learning. By employing a feature-prediction collaborative mechanism and gradient stopping techniques, the method significantly enhances the model's generalization capability for new category features while also mitigating the forgetting of old category knowledge during the incremental learning process to some extent. Experimental results demonstrate that the proposed method maintains detection performance for old categories throughout multiple incremental learning phases while simultaneously improving the learning capacity for new categories.