Abstract: Pre-training by numerous image data has become defacto for robust 2D representations. In contrast, due to the expensive data processing ... models to extract the multi-view visual features ...