Beyond Prompt Degradation: Prototype-guided Dual-pool Prompting for Incremental Object Detection

A team of researchers led by Yaoteng Zhang has unveiled a breakthrough framework for Incremental Object Detection (IOD), a critical challenge where AI vision models must learn new object categories over time without catastrophically forgetting previous knowledge. Their paper, "Beyond Prompt Degradation: Prototype-guided Dual-pool Prompting for Incremental Object Detection," accepted to the prestigious CVPR 2026 conference, proposes a novel prompt-decoupled framework called PDP. This approach directly tackles the core limitations of existing prompt-based IOD methods, specifically prompt coupling and prompt drift, which degrade model performance during continual learning.

The PDP framework's innovation lies in its dual-pool prompting paradigm, which explicitly separates prompts into a shared pool for task-general knowledge and a private pool for task-specific features. This decoupling prevents interference and mitigates prompt degradation. Complementing this is a Prototypical Pseudo-Label Generation (PPG) module that dynamically updates class prototypes to maintain consistent supervisory signals, countering the drift that occurs when old objects are mislabeled as background in new tasks. The results are substantial, with PDP achieving a 9.2% average precision (AP) improvement on the challenging MS-COCO benchmark and a 3.3% AP gain on PASCAL VOC, setting a new state-of-the-art. This work represents a significant step toward more stable and plastic vision systems capable of lifelong learning, with open-sourced code facilitating further research and application.