Cross-group bidirectional fusion network for 3D model classification

3D model classification is an important task in artificial intelligence, which aims to categorize 3D models based on shapes, structures, and other features. Focusing on the issues of insufficient inter-view interaction and inefficient feature fusion in some existing multi-view 3D model classification methods, this paper proposes a 3D model classification method based on Cross-group bidirectional fusion network. Firstly, dual-branch shared Swin Transformer feature extractor is combined with feature pyramid network to achieve multi-scale hierarchical feature fusion. Secondly, the cross-group interaction module constructs an attention-based semantic dependency matrix between the two view groups, enabling context-conditioned cross-view feature association. Thirdly, bidirectional feature-wise linear modulation strategy is proposed to adaptively calibrate cross-view feature responses by dynamically generating affine transformation parameters using the global semantics of the other view group. Fourthly, a stochastic hybrid data augmentation strategy is adopted during training to enhance regularization. Experimental results show that the proposed method achieves a classification accuracy of 95.58% on the ModelNet40 dataset, verifying the effectiveness of the proposed mechanism in facilitating efficient multi-view feature fusion.