DANCE: Density-agnostic and Class-aware Network for Point Cloud Completion

TL;DR

Point cloud completion must cope with incomplete scans that vary in sparsity. DANCE addresses this by a density-agnostic, class-aware pipeline that generates candidate points $P^S$ via ray-based sampling with $M = V R^2$, refines them through a transformer decoder predicting offsets $o_m$ and opacities $\,\\sigma_m$, and yields a flexible output $P^{out}$ combined with the observed input $P^I$ to form $P^{pred}$. A lightweight 3D classification head provides semantic priors from the incomplete geometry, guiding completion without image supervision, while a fusion network translates features and priors into per-point predictions. Experiments on PCN and MVP show state-of-the-art accuracy, stronger structural consistency, and robustness to density and noise, with the added benefit that output density can be controlled at inference by adjusting the sampling parameter $R$.

Abstract

Point cloud completion aims to recover missing geometric structures from incomplete 3D scans, which often suffer from occlusions or limited sensor viewpoints. Existing methods typically assume fixed input/output densities or rely on image-based representations, making them less suitable for real-world scenarios with variable sparsity and limited supervision. In this paper, we introduce Density-agnostic and Class-aware Network (DANCE), a novel framework that completes only the missing regions while preserving the observed geometry. DANCE generates candidate points via ray-based sampling from multiple viewpoints. A transformer decoder then refines their positions and predicts opacity scores, which determine the validity of each point for inclusion in the final surface. To incorporate semantic guidance, a lightweight classification head is trained directly on geometric features, enabling category-consistent completion without external image supervision. Extensive experiments on the PCN and MVP benchmarks show that DANCE outperforms state-of-the-art methods in accuracy and structural consistency, while remaining robust to varying input densities and noise levels.

Figures (11)

• Figure 1: Limitations of previous methods. (a) Fixed input/output densities. Prior methods assume fixed-size inputs and outputs, which limits their applicability in real-world scenarios. (b) Image-based generative frameworks. Methods relying on 2D image representations often produce shapes that do not align with the original 3D input geometry.
• Figure 2: The overall pipeline of DANCE. It generates candidate points using ray-based sampling strategy to reconstruct missing regions of incomplete point clouds. The encoder extracts features from both the incomplete input and generated points, and the decoder processes them to estimate the offset and opacity of each point. Final completion results are obtained by refining candidate points using the predicted offset and selecting valid ones based on the predicted opacity. Best viewed in color.
• Figure 3: Candidate point generation in DANCE. Candidate points are generated using a ray-based sampling strategy. Colors represent the corresponding faces and their associated candidate points. Best viewed in color.
• Figure 4: Offset and opacity estimation in DANCE. The estimated offset and opacity refine the position of each candidate point and control its inclusion in the final completion.
• Figure 5: Architecture of the transformer in DANCE. Cross-attention between $f^S_v$ and $f^I$ followed by self-attention within viewpoint groups $v$.