PointMCF: Multi-Modal Conditioned Flow with Masked Point and Global Distribution for 3D Representation Learning

In the 3D self-supervised representation learning (3D SSRL) community, masked point modeling (MPM) and global distribution modeling (GDM) are dominant paradigms with strong generalization capability. MPM produces unevenly distributed occluded areas due to random masking, leading to large variations in reconstruction difficulty. GDM increases overall difficulty via noise perturbations, yet the difficulty distribution remains uniform, resulting in insufficient focus on critical regions. Moreover, these point-to-point generative methods fail to fully exploit complementary multi-modal information. To address these limitations, we propose PointMCF, a 3D SSRL framework that leverages both paradigms. Their fundamentally different optimization objectives make effective integration challenging. We model a Leap-step Vector Field that maps noise to data distributions to capture representation evolution. A Unified Modal Bridge adaptively incorporates image, text, and point cloud conditions, while a Coarse-to-Fine Grouping Module enhances hierarchical feature understanding. Experimental results demonstrate that PointMCF consistently achieves superior performance across multiple real-world datasets, excelling in object classification, few-shot and zero-shot learning, medical diagnosis, semantic segmentation, part segmentation, and 3D action recognition. Notably, on three variants of ScanObjectNN, PointMCF outperforms the multi-modal method ReCon by 1.26% and surpasses the baseline Point-MAE by 6.46%.