Abstract: Highlights•Designs an online open world object detection baseline.•Takes advantage of prototypes and statistics of old knowledge.•Proposes dual-level perturbations for capturing new and old knowledge.•Excellent results prove the promising performance of our method and learning manner.
Loading