Cross-Instance Contrast with Color Decay for Image-Based 3D Shape Retrieval

Image-Based 3D Shape Retrieval (IBSR) has lately attracted appealing interests due to the proliferation of 3D shapes and the maturation of 2D representation learning. However, most of the recent works focus on object shapes with less object color in 2D image representations, which cannot well capture the geometric characteristics. Moreover, the relations between multiple instances which correspond to the same 3D shape are neglected in existing methods, which limits retrieval performance. To tackle these challenges, we propose a Cross-Instance Contrast with Color Decay (CIC ² D) approach for IBSR. Specifically, a straightforward but effective color decay technique is designed to enlarge the color space, which is beneficial for investigating more geometric characteristics. Furthermore, a cross-instance contrastive loss is proposed to build relations between multiple instances for capturing robust 3D representations. Extensive experiments demonstrate that our CIC ² D outperforms advanced methods on three representative benchmark datasets.