CF3d: Category Fused 3D Point Cloud Retrieval

3D point cloud retrieval technology that facilitates resource reuse has become a hot research topic in the field of computer vision. In recent years, many view-based retrieval methods have been proposed. Despite achieving state-of-the-art performance in many benchmarks, these methods inevitably lose a large amount of spatial information due to the nature of the view projection process. In this paper, we propose a category-fused retrieval method that directly extracts geometric and semantic features from the 3D point cloud. Specifically, we incorporate category information by learning a separate network for point cloud classification. Apart from the conventional cross-entropy loss, we design an intra-class constrained loss function to make the intra-class features more compact. We design an offset-attention module with an implicit Laplacian operator to reduce the noise in our feature learning process. In addition, we devise a data-driven 3D augmentation module that learns to generate difficult but meaningful examples for model training. Consistency loss is added to ensure that the augmented sample lies close to its counterpart in the feature space. Extensive experiments are conducted on synthetic datasets (i.e. ModelNet40 and ShapeNetPart) and the real scanned dataset of ScanObjectNN to demonstrate that our method outperforms state-of-the-art methods.