EVENT DOSSIER
To address the performance degradation in cross-domain 3D category incremental learning due to the heterogeneity of point cloud data sources, researchers proposed the PolyMem method. This study constructed the Domain3D-CIL protocol, which includes various heterogeneous sources such as clean CAD domains, RGB-D camera scans, video reconstructions, and damaged observations, and verified that there are significant cross-domain performance differences between mainstream methods. As a method that does not require examples, PolyMem enhances cross-domain robustness by implicitly modeling high-order statistical information of feature distributions. Experiments show that this method can effectively mitigate performance differences while improving the overall performance of the model across different domains.
针对跨域 3D 类别增量学习（Cross-Domain 3D Class-Incremental Learning）中因点云数据源异构导致的性能差异问题，研究者提出了 PolyMem 方法。该研究首先建立了包含来自清洁 CAD 域、RGB-D 相机扫描、视频重建及受损观察等多种异构源的 Domain3D-CIL 训练与评估协议，并验证了主流 CIL 方法均存在显著的跨域性能下降现象。PolyMem 作为一种无需示例（exemplar-free）的方法，通过隐式建模特征分布的高阶统计信息来增强跨域鲁棒性。实验表明，该方法能有效缓解性能差异，同时提升模型在各领域的整体表现。