One-Shot Replay: Boosting Incremental Object Detection via Retrospecting One Object

Modern object detectors are ill-equipped to incrementally learn new emerging classes over time due the well-known phenomenon of catastrophic forgetting. Due data privacy or limited storage, few no images old can be stored for replay. In this paper, we design a novel One-Shot Replay (OSR) method incremental detection, which is an augmentation-based method. Rather than storing original images, only one object-level sample each class reduce memory usage significantly, and find that copy-paste harmonious way replay detection. learning procedure, diverse augmented samples with co-occurrence objects existing training generated. To introduce more variants classes, propose two augmentation modules. The module aims enhance ability detector perceive potential unknown objects. feature explores relations between augments space via analogy. Extensive experimental results on VOC2007 COCO demonstrate OSR outperform state-of-the-art detection methods without using extra wild data.