A Multimodal framework for 3D few-shot class-incremental learning

Zhang Senbao

doi:10.1007/s00530-025-01957-5

A Multimodal framework for 3D few-shot class-incremental learning

2025-09-25

0

PRE

AI

DOI：10.1007/s00530-025-01957-5

Original

Original request for help

Save

Abstract

En 中文

Few-shot class-incremental learning (FSCIL) is the challenge of learning new categories with limited labeled samples while preserving previously learned knowledge. Although existing methods mainly focus on 2D images, the increasing use of 3D data, combined with its inherent scarcity, necessitates exploring FSCIL for 3D data. Traditional approaches rely on projecting 3D point clouds to 2D planes, causing significant loss of geometric information. This loss occurs because 2D images cannot fully represent the 3D structure. To address this issue, we propose a multimodal framework that integrates the benefits of 2D depth maps and 3D point clouds for improved feature extraction. Additionally, we designed a Task-Adaptive Learning Strategy (TLS) to reduce catastrophic forgetting and enhance performance. TLS adjusts the model’s adaptability based on task-specific features, improving its incremental learning capacity. Experimental results demonstrate that our method outperforms state-of-the-art techniques across several standard 3D FSCIL benchmark datasets.

Keywords:

Few-shot class-incremental learning

Multimodal

3D point cloud

A Multimodal framework for 3D few-shot class-incremental learning

Abstract

Journal

Organization