Incremental Learning for Multi-view 3D Shape Recognition based on Knowledge Distillation

Abstract

Due to the convenience of image acquisition and processing, view-based 3D shape recognition methods have attracted much attention. The existing view-based methods focus on improving recognition accuracy but ignore the lightweight of model that is important for applications. In this paper, we propose a method for incremental learning of multi-view 3D shape recognition based on knowledge distillation. The proposed method trains models for new tasks only with incremental data and shares parameters for different recognition tasks, which can significantly reduce the computational and storage costs of the model. We use MVCNN as the original model to validate the effectiveness of our method on the ModelNet dataset. The experiments show that our method is very effective for incremental learning of 3D shape recognition.