CKCA-Det: class knowledge constraint-adaption for incremental few shot object detection

Few-Shot Object Detection (FSOD) requires models to generalize to novel classes under scarce training samples. Incremental Few-Shot Object Detection (iFSOD) further demands continuous learning of novel classes while preserving base classes knowledge and mitigating catastrophic forgetting. However, most existing FSOD methods suffer from severe catastrophic forgetting due to their reliance on random dot-product aggregation between class prototypes and query features, which neglects inter-class confusion and imbalance between base and novel categories. This approach fails to enhance few-shot generalization or retain prior knowledge effectively. To address these limitations, we propose Class Knowledge Constraint-Adaption (CKCA)-Det, a novel framework for iFSOD. Specifically, we design a Class-Adaptive RoI-Head that employs a multi-head attention mechanism to adaptively aggregate class prototypes most relevant to the query set, thereby enhancing query feature representations. Additionally, we introduce a cosine-normalized classifier and propose an Old-Knowledge Constraint (OKC) to regularize the fine-tuned parameters of base classes, significantly reducing knowledge forgetting. Experimental results on public datasets demonstrate the excellent performance obtained by our approach.