DANCE: Density-Agnostic and Class-Aware Network for Point Cloud Completion

Conference: AAAI2026
arXiv: 2511.07978
Code: ayeong0909/DANCE
Area: 3D Vision
Keywords: point cloud completion, density-agnostic, class-aware, transformer, ray-based sampling, opacity prediction

TL;DR

The proposed DANCE framework achieves density-agnostic point cloud completion through ray-based candidate point sampling and an opacity prediction mechanism, while introducing a classification head to provide semantic priors, achieving state-of-the-art performance on PCN and MVP benchmarks.

Background & Motivation

Point cloud completion aims to recover missing geometric structures from incomplete 3D scans caused by occlusion or sensor perspective limitations. It constitutes a key preprocessing task for autonomous driving, robotics, and 3D reconstruction.

Existing methods have two core limitations:

1. Fixed Density Assumption: The vast majority of methods assume a fixed density for input and output point clouds (e.g., outputting a fixed number of 4096 points), which fails to adapt to sparsity variations caused by different object distances and sensor resolutions in real-world scenes.
2. Reliance on Image Supervision: Recent generative methods (such as GenPC, PCDreamer) convert partial point clouds into 2D images and then utilize image-to-3D models for completion. Such strong 2D priors often lead the completion results to deviate from the original 3D geometry.

An ideal completion method should: (a) be density-agnostic, handling inputs of arbitrary sparsity and flexibly controlling the output density; (b) directly learn semantic priors from 3D geometry, rather than relying on external image representations.

Core Problem

How to complete only the missing regions and preserve the observed geometry without relying on fixed density and image supervision, while introducing class semantic information to enhance completion quality?

Method

Overall Architecture

DANCE consists of three stages: candidate point generation -> encoder feature extraction -> decoder completion prediction.

1. Ray-based Sampling

Inspired by NeRF, \(V\) viewpoints are placed around the incomplete point cloud (default \(V=6\), forming a hexahedron), where each viewpoint corresponds to a face facing the object. An \(R \times R\) grid is placed on each face (default \(R=21\)). Rays are cast from the viewpoints through each grid point, and a 3D candidate point is sampled along the ray based on a Gaussian distribution. This generates a total of \(M = V \cdot R^2\) candidate points \(P^S\).

These candidate points are initially coarse and are subsequently refined into precise locations by the encoder-decoder.

2. Encoder (3D Feature Extraction)

The candidate points \(P^S\) and the incomplete point cloud \(P^I\) share the same encoder \(E\) (which can be PointNet, DGCNN, etc.) to extract:

• Candidate features \(f^S = E(P^S) \in \mathbb{R}^{M \times d_{en}}\)
• Global features \(f^I = \text{maxpool}(E(P^I)) \in \mathbb{R}^{1 \times d_{en}}\)

The shared encoder ensures that both sets of features are aligned within the same feature space.

3. Decoder

(a) Face Transformer: Candidate features are processed in groups according to their viewpoints. Each viewpoint group \(f_v^S\) first performs cross-attention with the global feature \(f^I\) (injecting a global shape prior), and then performs self-attention within the group (enhancing local geometric consistency). Viewpoint positional encodings \(E_v^{fpos}\) are utilized to maintain spatial relationships.

(b) Classification Head: Predicting class probability distributions \(\mathbf{p}^{cls} \in \mathbb{R}^c\) on the global feature \(f^I\) via an MLP + softmax, providing semantic priors for completion.

(c) Fusion Network: Geometric features \(F^S\) are first passed through a compression-expansion MLP (with a bottleneck dimension of 4, aligned with the output dimension) and then concatenated with class probabilities \(\mathbf{p}^{cls}\). The prediction head then outputs the following for each candidate point:

• Offset \(o_m = \{o_x, o_y, o_z\}\): Position correction in the local coordinate system with the candidate point as the origin.
• Opacity \(\sigma_m\): Determines whether the point is valid (retained if \(\sigma \geq 0.5\)).

The final missing point cloud is \(P^{out} = \{p_m + o_m \mid \sigma_m \geq 0.5\}\), which is merged with the input to obtain \(P^{pred} = P^I \cup P^{out}\).

4. Loss & Training

\[\mathcal{L}_{total} = \lambda \cdot \text{CD}(P^{pred}, P^{GT}) + (1-\lambda) \cdot \mathcal{L}_{cls}\]

where CD represents Chamfer Distance, and \(\mathcal{L}_{cls}\) is the cross-entropy classification loss.

Key Experimental Results

Main Results on PCN Dataset (8 Classes, L1-CD)

Method	CD-Avg ↓	DCD-Avg ↓	F1 ↑
SVDFormer	6.61	0.534	0.848
CRA-PCN	6.56	0.537	0.846
PCDreamer	6.52	0.531	0.856
DANCE (Ours)	6.46	0.528	0.859

Main Results on MVP Dataset (16 Classes)

Resolution	CD-Avg ↓	F1 ↑
4096 points	4.19	0.662
8192 points	3.37	0.754

Both outperform previous SOTA models such as DualGenerator and PDR.

Ablation Study (PCN)

• Removing Classification Head: CD-Avg increases from 6.42 → 6.46, F1 decreases from 0.859 → 0.856.
• Removing Face Attention: CD-Avg increases from 6.42 → 6.52, F1 decreases from 0.859 → 0.849.

Both components contribute positively, with the Face Transformer having a greater impact.

Robustness

• Noise Robustness: Under different Gaussian noise levels, the performance degradation of DANCE is smaller than that of SVDFormer and SeedFormer.
• Density Flexibility: While \(R=21\) is fixed during training, \(R\) can be directly changed to \(R=17\) or \(R=29\) during inference to adjust the output density without retraining.

Highlights & Insights

1. Density-Agnostic Design: This work achieves density-agnostic 3D point cloud completion for the first time, where both input and output densities are variable, and the number of output points is naturally controlled through opacity filtering.
2. Pure 3D Semantic Priors: The classification head directly learns class information from 3D geometric features without requiring image supervision, making it more suitable for real-world deployment.
3. Completing Only Missing Regions: The original observed geometry is preserved, avoiding detail loss caused by global regeneration.
4. Ray Sampling Strategy: The NeRF concept is elegantly transferred to point cloud completion, providing a structured distribution of candidate points.
5. Controllable Density at Inference: Adjusting \(R\) allows flexible changes to the output resolution, offering high practicality.

Limitations & Future Work

1. Fixed Viewpoint Configuration: Placing fixed hexahedral viewpoints and uniform grids may not be the optimal sampling strategy for highly asymmetric or complex structured objects.
2. Candidate Point Overhead: \(M = V \cdot R^2\) yields a large number of candidate points when \(R\) is large, increasing computational overhead.
3. Limited Classes: The classification head relies on a predefined set of classes, which raises issues regarding generalization capability to unseen classes in the training set.
4. Evaluation Only on Synthetic Data: Both PCN and MVP are based on synthetic data from ShapeNet, and have not been validated on real sensor scan data (such as ScanNet, KITTI).
5. Future direction proposed by the authors: Adaptive viewpoint sampling, dynamically adjusting sampling positions and the number of viewpoints according to the input geometric features.

Related Work & Insights

Dimension	PCN / PoinTr	GenPC / PCDreamer	DANCE
Completion Scope	Global / Missing only	Global Regeneration	Missing regions only
Density Assumption	Fixed input/output	Fixed	Density-Agnostic
Semantic Prior	None	2D image supervision	3D classification head
Output Density Control	Unsupported	Unsupported	Adjustable \(R\) at inference

While sharing the "missing only" paradigm with PoinTr, DANCE achieves density flexibility through the opacity mechanism. Furthermore, with the introduction of semantic priors, the CD-Avg of DANCE on PCN decreases to 6.46 compared to PoinTr's 7.76.

Inspirations & Connections

• Cross-domain Transfer from NeRF to Point Cloud Completion: The concept of ray sampling + opacity prediction can be extended to other 3D generation tasks (e.g., scene completion, point cloud upsampling).
• Lightweight Semantic Guidance: A simple classification head significantly improves completion quality, suggesting that semantic priors can be introduced cost-effectively in other 3D tasks.
• Density-Controllable Inference: The opacity filtering mechanism can be adapted to scenarios requiring flexible control over output resolution (e.g., Level of Detail (LOD) generation).

Rating

• Novelty: ⭐⭐⭐⭐ — The density-agnostic design and NeRF-style sampling transfer are novel, although the classification head design is relatively simple.
• Experimental Thoroughness: ⭐⭐⭐⭐ — Comprehensive baseline comparisons on PCN/MVP and complete ablation studies are provided, but real-world data validation is lacking.
• Writing Quality: ⭐⭐⭐⭐ — The motivation is clear, the structure is well-organized, and the illustrations aid understanding effectively.
• Value: ⭐⭐⭐⭐ — The direction of density-agnostic completion holds practical significance and provides valuable insights for real-world deployment.

Related Papers

• [AAAI 2026] Rethinking Multimodal Point Cloud Completion: A Completion-by-Correction Perspective
• [AAAI 2026] DAPointMamba: Domain Adaptive Point Mamba for Point Cloud Completion
• [AAAI 2026] Class-Partitioned VQ-VAE and Latent Flow Matching for Point Cloud Scene Generation
• [ECCV 2024] Explicitly Guided Information Interaction Network for Cross-modal Point Cloud Completion
• [CVPR 2026] Geometric-Aware Hypergraph Reasoning for Novel Class Discovery in Point Cloud Segmentation