3D continual learning encounters significant challenges such as catastrophic forgetting, task interference, and semantic confusion, primarily due to the sparse structure and high variability of point cloud data. To address these issues, we propose StarK-3D (Semantic-aware Replay and Key knowledge fusion for 3D Continual Learning), which consists of two novel components: Semantic-Aware Replay (SAR) and Key Knowledge Fusion (KKF). Specifically, SAR selectively generates and replays the more confusing samples by focusing on the semantic similarity between categories, enabling targeted review without excessive memory overhead. KKF leverages gradient information to indicate the importance of parameters and integrates the indicated key knowledge into the updating direction, enabling effective fusion of task-relevant knowledge during training. We conduct comprehensive experiments across task-incremental, domain-incremental, and scene-incremental settings. The results show that StarK-3D achieves state-of-the-art performance, demonstrating strong effectiveness and robustness in 3D continual learning.
Qian et al. (Sun,) studied this question.