ContinualCropBank: Object-Level Replay for Semi-supervised Online Continual Object Detection

Object detection models typically rely on static, fully labeled datasets— an unrealistic assumption for dynamic environments. Continual Learning (CL) counters this by enabling models to learn new tasks without forgetting previous ones, yet most methods require known task boundaries and multiple data passes. Online Continual Learning (OCL) is more practical, as it processes data in a single pass, but it relies heavily on costly annotations. To address this limitation, Label-Efficient Online Continual Object Detection (LEOCOD) extends OCL with a semi-supervised formulation that leverages both unlabeled and limited labeled data. In this paper, we propose ContinualCropBank, an object-level replay module for LEOCOD that stores cropped object patches and pastes them into training images, providing fine-grained replay. This solution mitigates catastrophic forgetting, alleviates foreground–background imbalance, and increases exposure to small objects. Experiments on two benchmark datasets demonstrate that incorporating ContinualCropBank improves detection accuracy and resilience to forgetting, achieving gains of up to 9.57 percentage points in average accuracy and reducing degradation from forgetting by up to 2.32 points. Our code is available at https://github.com/lucastsutsui/ContinualCropBank .