Rotation-Invariant Completion Network

TL;DR

This work tackles the problem of completing incomplete point clouds under arbitrary poses by introducing Rotation-Invariant Completion Network (RICNet).RICNet combines a Rotation-Invariant Embedding Module with a Dual Pipeline Completion Network (DPCNet) and an Enhancing Module (RENet) to produce coarse and refined complete point clouds, utilizing a shared-weight, dual-path VAE and KL-based consistency losses between latent distributions.The method explicitly leverages rotation-invariant global features (via RIConv++) and local features (via DGCNN) to maintain consistent representations despite rigid transformations, and it employs a refinement stage to recover fine-grained structure and relational cues.Evaluations on the MVP dataset with random rotations show that RICNet outperforms existing approaches in both original and rotated settings, with ablations confirming the essential roles of the rotation-invariant encoder, dual-path design, and enhancement module, making it robust for real-world applications like robotics and autonomous driving.

Abstract

Real-world point clouds usually suffer from incompleteness and display different poses. While current point cloud completion methods excel in reproducing complete point clouds with consistent poses as seen in the training set, their performance tends to be unsatisfactory when handling point clouds with diverse poses. We propose a network named Rotation-Invariant Completion Network (RICNet), which consists of two parts: a Dual Pipeline Completion Network (DPCNet) and an enhancing module. Firstly, DPCNet generates a coarse complete point cloud. The feature extraction module of DPCNet can extract consistent features, no matter if the input point cloud has undergone rotation or translation. Subsequently, the enhancing module refines the fine-grained details of the final generated point cloud. RICNet achieves better rotation invariance in feature extraction and incorporates structural relationships in man-made objects. To assess the performance of RICNet and existing methods on point clouds with various poses, we applied random transformations to the point clouds in the MVP dataset and conducted experiments on them. Our experiments demonstrate that RICNet exhibits superior completion performance compared to existing methods.

Figures (3)

• Figure 1: The architecture of RICNet. DPCNet generates a coarse complete point cloud ${Y_{c}}'$, while the enhancing module refines the fine-grained details of the final output to produce the final completion ${Y_{f}}'$. DPCNet employs an encoder-decoder architecture, with the two parallel paths sharing weights. The distribution link ensures consistency between the posterior distribution of the encoded partial point cloud and the prior distribution of the complete point cloud.
• Figure 2: A Rotation-Invariant Convolutional Layer. We perform farthest point sampling to select a set of reference points $\left \{ r_{j} \right \}$ (red dots) from the point cloud. K-NN is utilized to obtain local point sets $X_{j}=\left \{x_{n}\right \}$ (blue dots) for each reference point $r_j$. IRIFs are then calculated and transformed into a high-dimensional space using an MLP. These features are combined with previous layer features (if any), followed by pointwise convolution and Maxpooling to generate the New Features.
• Figure 3: IRIF Construction. Given a reference point r, the K-NN algorithm retrieves k nearest neighbors $X_{j}=\left \{x_{n}\right \}$ and orders the points clockwise. The IRIF of a neighbor point $x_{n}$ is represented by (\ref{['eq2']}).