Cooperative meta-learning for incremental few-shot object detection in open urban environments

Incremental few-shot object detection (iFSOD) plays a pivotal role in enabling spatially embodied intelligence in unmanned systems which operate in open urban environments with novel objects and limited annotations. For example, to ensure robust decision-making, autonomous agents must continually expand their perception capabilities while preserving previously acquired knowledge. However, existing iFSOD approaches are built either on two-stage or DETR-based architectures. These approaches are inept to balance stability and plasticity when they are adapting to new object categories. This limitation stems from the lack of principled mechanisms that support continual adaptation across tasks. To address this challenge, in this paper we propose a meta-learning-based incremental DETR framework that disentangles task-specific adaptation and transferable meta-knowledge through a cooperative design. In specific, a task learner performs fine-grained adaptation to novel objects by fusing multi-scale support and query features. Meanwhile, a meta-learner captures cross-task transferable representations to enhance generalization and mitigate catastrophic forgetting. The proposed cooperative scheme enables stable continual learning from limited data while maintaining knowledge of previously learned classes. Experiments on MS COCO and COCO2VOC demonstrate superior performance of our method in both detection accuracy and base–novel balance.