MF-Adapter: Better 3D Foundation Model with Multimodal Fusion Adapter

Recent advancements in 3D vision foundation models have significantly enhanced the generalization ability of 3D representation learning. However, existing models primarily rely on contrastive learning, limiting their capacity to fully exploit multimodal information for downstream tasks. To address this, we propose MF-Adapter, a multimodal fusion adapter designed to improve the adaptability of 3D foundation models. Our method introduces a multimodal fusion mechanism, including self-attention enhancement, cross-modal correlation, and residual feature fusion, enabling a more effective integration of point cloud and multi-view features. Additionally, we design implicit supervision signals based on multimodal consistency loss and dynamic pseudo-label loss to enhance model robustness without requiring labeled data. Extensive experiments on ModelNet40, ModelNet10, and ScanObjectNN demonstrate that our approach achieves state-of-the-art performance.