Congratulations to Jianzhou Wang on IJCAI 2025 (CCF - A) Paper Acceptance

Diffuse&Refine: Intrinsic Knowledge Generation and Aggregation for Incremental Object Detection IJCAI2025 （CCF-A）

Incremental object detection (IOD) aims to gradually enhance the ability of an object detector to recognize new categories. However, the representational confusion between old and new categories leads to the problem of catastrophic forgetting. To alleviate this problem, we propose the DiffKA method, which generates and aggregates intrinsic knowledge through forward diffusion and reverse diffusion, and gradually constructs rigid category boundaries. Specifically, forward diffusion generates potential inter-class correlations in the hierarchical tree structure (referred to as the Intrinsic Correlation Tree, ICT) through information propagation. Subsequently, reverse diffusion aggregates and refines the semantic features of the tree nodes, explicitly establishing rigid category boundaries in the semantic space. To maintain semantic consistency, we restructure the semantic relevance of knowledge through a paradigm, enabling the adaptive and dynamic update of the Intrinsic Correlation Tree. Experiments on the MS COCO dataset demonstrate that DiffKA achieves state-of-the-art performance in the incremental object detection task, showcasing significant advantages.