SPAC-Net: Rethinking Point Cloud Completion with Structural Prior

TL;DR

This paper proposes a novel framework, termed SPAC-Net, that aims to rethink the completion task under the guidance of a new structural prior, called interface, and predicts the coarse shape by learning the displacement from the points in interface move to their corresponding position in missing parts.

Abstract

Point cloud completion aims to infer a complete shape from its partial observation. Many approaches utilize a pure encoderdecoder paradigm in which complete shape can be directly predicted by shape priors learned from partial scans, however, these methods suffer from the loss of details inevitably due to the feature abstraction issues. In this paper, we propose a novel framework,termed SPAC-Net, that aims to rethink the completion task under the guidance of a new structural prior, we call it interface. Specifically, our method first investigates Marginal Detector (MAD) module to localize the interface, defined as the intersection between the known observation and the missing parts. Based on the interface, our method predicts the coarse shape by learning the displacement from the points in interface move to their corresponding position in missing parts. Furthermore, we devise an additional Structure Supplement(SSP) module before the upsampling stage to enhance the structural details of the coarse shape, enabling the upsampling module to focus more on the upsampling task. Extensive experiments have been conducted on several challenging benchmarks, and the results demonstrate that our method outperforms existing state-of-the-art approaches.

Figures (9)

• Figure 1: Our method is presented in a comparative visualization with other state-of-the-art (SOTA) point cloud completion techniques. In the comparative figures, two detail areas from the completion targets are selected to illustrate the effectiveness of the different approaches. The cyan points denote the predicted point cloud, whereas the blue points represent the input partial scans.
• Figure 2: (a) shows several visualization results of interface(yellow points) in incomplete object. $\textbf{(b)}$ illustrates the interface in partial scans is equivalent to the one in missing parts, which means that $\textit{interface}_{1}$ equals to $\textit{interface}_{2}$, $\textit{interface}_{3}$ equals to $\textit{interface}_{4}$. So we can localize interface in partial scans to establish spatial perception on missing parts before predicting them.
• Figure 3: The overall framework of SPAC-Net. Given an input $P$, MAD module is used to localize the interface and the Encoder is applied to extract the feature according to $P$. During shape prediction, a coarse shape is generated by learning the displacement from the interface to the missing part. To further facilitate the recovery of details, we leverage several stacked SSP modules to augment structure details before shape upsampling. Finally, we restore missing parts at high resolution by an upsampling module, and add it with input $P$ to obtain the complete point cloud. In the figure, $N_T$ is the number of interface, $C_P$, $C_M$ and $C_T$ are the dimension of corresponding feature, respectively.
• Figure 4: Interface Localization Methods Illustration. (a) and (b) depict the methods under the scenarios of occluded points available and occluded points unknown, respectively.
• Figure 5: Qualitative comparison of SPAC-Net and state-of-art methods on ShapeNet-55 dataset, showing Airplane, Printer, Chair and Bench from top to bottom.The blue points indicate partial scans, the cyan points signify the predicted results, the magenta points denote the interfaces, and the brown points correspond to the ground truth.