Multi-modality integrated class incremental learning networks for 3D object recognition

Yufei Zhang

doi:10.1016/j.knosys.2025.114778

Multi-modality integrated class incremental learning networks for 3D object recognition

2025-11-02

0

PRE

AI

DOI：10.1016/j.knosys.2025.114778

Original

Original request for help

Save

Abstract

En 中文

Multimodal three-dimensional (3D) object recognition is garnering attention across various fields owing to its robust performance and wide application potential. To enable multimodal learning networks such that new data and tasks can be adapted continually in 3D object recognition, we introduce a novel method named multimodal integration class incremental learning (MICIL). MICIL simplifies nonlinear network learning into linear segment learning, making model more adaptable to continual learning tasks. This approach enables MICIL to learn new tasks without forgetting them, thus resulting in high performance. As MICIL does not require the retraining or reviewing of old tasks when learning new ones, it reduces memory occupancy and ensures data privacy. Additionally, in the proposed MICIL, an output-integration module is incorporated to combine the outputs from the feature extractors and fusion network, thereby enhancing the robustness of the network. To the best of our knowledge, this is the first evaluation protocol used in multimodal 3D object classification that focuses on measuring network performance during long learning phases. Experimental results show that the proposed method achieves state-of-the-art performance in all evaluation tasks.

Multi-modality integrated class incremental learning networks for 3D object recognition

Abstract

Journal

Organization