Contrastive R-CNN for Incremental Learning in Object Detection

Incremental learning for image classification has been widely studied in the past few years, but few works explored incremental learning for object detection. Most existing incremental object detectors deploy knowledge distillation to constrain the model to retain old knowledge and avoid the catastrophic forgetting phenomenon. Nonetheless, this common practice results in strong constraints adhering to the old knowledge, therefore deteriorating the learning ability to new knowledge. In this work, we propose a new framework named Contrastive R-CNN for incremental learning of object detection to balance the retaining of the old knowledge and the learning of the new knowledge. The proposed framework is mainly composed of two modules, data distillation and temporal contrast. Data distillation presents a median entropy filter strategy to generate the annotations for the RoIs of the old objects, while temporal contrast designs an RoI contrast mechanism to minimize the ambiguity between old and new instances for better incremental learning. Extensive experiments on the PASCAL VOC dataset demonstrate the effectiveness of our proposed approach.