DIODE: Dilatable Incremental Object Detection

To accommodate rapid changes in the real world, cognition system of humans is capable continually learning concepts. On contrary, conventional deep models lack this capability preserving previously learned knowledge. When a neural network fine-tuned to learn new tasks, its performance on trained tasks will significantly deteriorate. Many recent works incremental object detection tackle problem by introducing advanced regularization. Although these methods have shown promising results, benefits are often short-lived after first step. Under multi-step learning, trade-off between old knowledge and task becomes progressively more severe. Thus, regularization-based detectors gradually decays for subsequent steps. In paper, we aim alleviate decay proposing dilatable detector (DIODE). For task-shared parameters, our method adaptively penalizes important weights previous tasks. At same time, structure model dilated or expanded limited number task-specific parameters promote learning. Extensive experiments PASCAL VOC COCO datasets demonstrate substantial improvements over state-of-the-art methods. Notably, compared with methods, achieves up 6.0% improvement increasing just 1.2% each newly task.