DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization

TL;DR

DC-PCN tackles sampling variability in incomplete point clouds by learning a consistent representation via a dual-codebook quantization scheme. It employs a transformer-based encoder–decoder with an encoder-codebook $\\mathbf{C}_E$ and a decoder-codebook $\\mathbf{C}_D$, plus a quantized information exchanging (QIE) mechanism to fuse shallow and deep level features for improved completion. The approach uses region-based modeling and Chamfer-distance supervision on both coarse and complete outputs, plus a dedicated codebook loss that combines internal and external terms. Experiments on PCN, ShapeNet_Part, ShapeNet34, and KITTI demonstrate state-of-the-art or competitive results, with ablations validating the contributions of dual-codebooks and QIE and highlighting robustness to sampling variability and real-world data.

Abstract

Point cloud completion aims to reconstruct complete 3D shapes from partial 3D point clouds. With advancements in deep learning techniques, various methods for point cloud completion have been developed. Despite achieving encouraging results, a significant issue remains: these methods often overlook the variability in point clouds sampled from a single 3D object surface. This variability can lead to ambiguity and hinder the achievement of more precise completion results. Therefore, in this study, we introduce a novel point cloud completion network, namely Dual-Codebook Point Completion Network (DC-PCN), following an encder-decoder pipeline. The primary objective of DC-PCN is to formulate a singular representation of sampled point clouds originating from the same 3D surface. DC-PCN introduces a dual-codebook design to quantize point-cloud representations from a multilevel perspective. It consists of an encoder-codebook and a decoder-codebook, designed to capture distinct point cloud patterns at shallow and deep levels. Additionally, to enhance the information flow between these two codebooks, we devise an information exchange mechanism. This approach ensures that crucial features and patterns from both shallow and deep levels are effectively utilized for completion. Extensive experiments on the PCN, ShapeNet\_Part, and ShapeNet34 datasets demonstrate the state-of-the-art performance of our method.

Figures (5)

• Figure 1: Comparison between different approaches: (a) point-based approach and (b) ours.
• Figure 2: DC-PCN architecture. To achieve consistent and unambiguous latent representations for point clouds, a dual-codebook design is introduced to quantize features in both shallow and deep level, which promotes to obtain completions in high-quality.
• Figure 3: Visualization of codebook distributions of (a) the encoder-codebook and (b) and the decoder-codebook, where an obvious discrepancy between them can be observed.
• Figure 4: Visualization of the results on PCN. The results of our method show higher-level noise suppression and have refined geometric structure compared to those obtained from existing methods.
• Figure 5: Visualization of the results on ablation settings A, B, D and E. Our method exhibits superiority when dealing with fine-grained completion structures.