A Two-Stage Framework for Rotation-Robust 3D Object Classification

In this work, we aim at developing a rotation-robust framework for 3D object classification. To this end, we propose a novel two-stage framework that utilizes multi-view images to represent 3D objects. The first stage focuses on precise feature extraction from individual images. The second stage performs multi-view feature fusion and classification through a novel attention module and a two-layer pooling operation, which significantly enhance the model's robustness to rotation of objects. Experimental results on the ModelNet40-rotated dataset demonstrate that the proposed method achieves state-of-the-art performance, with significant improvements in instance accuracy and weighted class accuracy under various rotation settings, outperforming existing view-based methods.