MV-ViT: High-Quality Features and Simple Fusion for Superior Multi-View 3D Object Classification

We present MV-ViT, a novel and efficient approach for multi-view 3D object classification. MV-ViT leverages the power of pre-trained Vision Transformers (ViTs) for robust single-view feature extraction, coupled with a two-stage training strategy that enhances performance while minimizing computational overhead. Unlike methods relying on complex multi-view fusion mechanisms, we demonstrate that high-quality features extracted by a fine-tuned ViT, when aggregated via simple mean pooling, achieve state-of-the-art results. MV-ViT achieves 97.23% accuracy with 12 views on the ModelNet40 benchmark, surpassing methods employing significantly more elaborate fusion architectures. Ablation studies highlight the critical role of strong ViT-derived features in achieving superior performance, making complex fusion redundant.