Conflicts between Likelihood and Knowledge Distillation in Task Incremental Learning for 3D Object Detection

Abstract

In autonomous driving scenarios, edge cases require perception algorithms, like 3D object detection, to incrementally learn new data during a long term. To achieve it, previous methods seek help from knowledge distillation and recursively transfer knowledge from old models to new models. However, conflicts exist between the likelihood term and the distillation regularizer on both old and new knowledge. In this paper, we discuss the drawback of knowledge distillation in the task-incremental-learning scenario for 3D object detection and propose a New-Task-Aware Biased Sampling and Knowledge-Distillation-Aware Detection Loss to solve the conflicts. Based on the KITTI dataset, we thoroughly evaluate our proposed method from the aspects of both forward and backward transfer in the task incremental-learning scenario. A great margin of improvement on the whole task sequence (5.6 mAP) demonstrates the effectiveness of our proposed method.