TNT-Net: Point Cloud Completion by Transformer in Transformer

Estimating the overall structure of a point cloud from a partial 3D point cloud input is a crucial task in computer vision. However, existing point cloud completion methods often overlook object detail information and the local correlation within the incomplete point cloud. To address this challenge, we propose an enhanced point cloud completion approach called TNT-Net, which leverages a transformer in transformer architecture for accurate and refined point cloud completion. TNT-Net incorporates a local feature extraction module to capture long-range correlations within the input point cloud. Moreover, we introduce stacked feature extractors to simplify subsequent calculations and gather more comprehensive feature information on the spatial distribution of the point cloud. Additionally, we present an efficient method that integrates the kNN-transformer into the existing point transformer to address the deficiency of local detail information in previous works. This method enables TNT-Net to capture fine-grained object details and local correlations more effectively. Extensive experiments conducted on synthetic datasets PCN, ShapeNet55/34, and real-world dataset KITTI demonstrate the superior quantitative and qualitative performance of TNT-Net compared to state-of-the-art methods.