3DPAN-CIL: a prototype assisted network of class-incremental learning for 3D point clouds

ICLR 2026 Conference Submission19456 Authors

Keywords: class-incremental learning, point cloud classification, optimal transport

TL;DR: a prototype assisted network of class-incremental learning for 3D point clouds

Abstract: In response to the continuous influx of 3D point clouds encountered in practical training scenarios, we propose a novel incremental learning and classification approach designated as 3DPAN-CIL, specifically tailored for 3D point cloud data. This method initially establishes the 3D category prototype that encapsulates the feature embedding of point clouds within a latent space. Then, we wisely construct an optimal transport strategy on this prototype space for the migration of 3D category prototypes. This alignment ensures that the distribution of new category prototypes adheres as closely as possible to the relative spatial distribution of old category prototypes, significantly reducing the catastrophic forgetting in the training model. Additionally, to tackle the challenge of imbalanced old and new samples, we introduce a prior-guided knowledge distillation strategy aimed at addressing the model’s preference for new knowledge. We conduct a series of experimental evaluations on both synthetic datasets and real scanning datasets, demonstrating that our method surpasses existing state-of-the-art approaches in terms of average accuracy and average forgetting rate. Notably, in the context of average scene partitioning, our method achieves improvements of 4.5% in average accuracy and 1.47% in average forgetting rate compared to other top-performing methods. The model and code are available at: https://github.com/FlRiver/3DPAN-CIL.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 19456