Multi-View Multi-Scale Network for 3D Object Recognition and Retrieval

Multi-view based 3D object recognition methods have become a research hotspot in the field of 3D object recognition, which represent and recognize 3D objects by fusing multiple views from different angles. However, existing methods often overlook the complex spatial interactions between features when fusing multi-view features, while some important features or details may be lost, consequently affecting the effectiveness of 3D object recognition and retrieval. To address this issue, this paper proposes a Multi-View Multi-Scale Network (MVMSN) for 3D object recognition and retrieval. The proposed network enhances spatial feature interactions of view information through recursive gated convolution based on multi-scale features from different views, and fully integrates local shallow features with global features using a cross-attention mechanism. Through this multis-cale fusion approach, the proposed network can further explore correlations between views and obtain more comprehensive multi-view information, thereby improving the expressive capability of multi-view representations. Comparative experiments on widely used ModelNet40 and ModelNet10 datasets demonstrate the superior performance of our method, validating its effectiveness.