DAPP: double adversarial point cloud prediction

The completion of 3D point clouds is crucial for scene understanding in depth-camera-based sensing systems, especially as the accuracy of depth cameras increases, some scenes have begun to utilize depth cameras to complete simple measurement tasks. However, raw 3D point cloud data is inevitably incomplete, and the currently existing generative data sets are not enough to meet the imaging requirements of simulated depth cameras. In this paper, we introduce what we believe to be a new approach to reconstruct the missing parts of an incomplete point cloud. The DAPP (double adversarial point-cloud prediction) network employs two sets of adversarial neural network structures. Sharing the same generator, the two discriminators in DAPP work together to guide the point cloud generation in different resolutions. We further propose a feature distribution block with improved 3D Point transformers to understand the geometric shapes, and a 3D expansion unit to bring diversity in the generation of precise structures. We also employ PAConv blocks with the Euclidean distances between 3D coordinates in the network to guide the generation of point clouds with different densities. At the same time, multi-dimensional Fourier transforms are used in multiple places on the network to remind high-dimensional features of the overall point cloud shape where they are located. Combining the above modules, the DAPP learns 3D shapes well and is able to understand long-term semantic geometry relationships. With the discriminators designed differently, the network can focus on different features of the input point cloud in the coarse prediction and fine prediction phases. Furthermore, we introduce a new dataset called MultiView57, which better simulates real-world outputs from a single depth camera, as we show in Dataset 1. Quantitative and visual results show that DAPP outperforms existing methods on our dataset and commonly used ones.