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

Conference: AAAI 2026 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

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

Background & Motivation

Point cloud completion aims to recover missing geometric structures from incomplete 3D scans caused by occlusion or sensor viewpoint limitations, serving as a critical prerequisite for autonomous driving, robotics, and 3D reconstruction.

Existing methods suffer from two core limitations:

1. Fixed-density assumption: The vast majority of methods assume fixed densities for both input and output point clouds (e.g., a fixed output of 4096 points), making them ill-suited for real-world scenarios where sparsity varies with object distance and sensor resolution.
2. Reliance on image supervision: Recent generative methods (e.g., GenPC, PCDreamer) convert partial point clouds into 2D images and leverage image-to-3D models for completion; strong 2D priors often cause the completed results to deviate from the original 3D geometry.

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

Core Problem

How to complete only the missing regions while preserving observed geometry—without relying on fixed density assumptions or image supervision—and simultaneously incorporate class-level semantic information to improve completion quality?

Method

Overall Architecture

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

1. Candidate Point Generation (Ray-based Sampling)

Inspired by NeRF, \(V\) viewpoints (default \(V=6\), forming a hexahedral configuration) are placed around the incomplete point cloud, each corresponding to a face oriented toward the object. An \(R \times R\) grid (default \(R=21\)) is placed on each face, and rays are cast from each viewpoint through every grid point. One 3D candidate point is sampled along each ray according to a Gaussian distribution, yielding a total of \(M = V \cdot R^2\) candidate points \(P^S\).

These candidate points are intentionally imprecise and are subsequently refined to accurate positions 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 \(E\) (e.g., PointNet, DGCNN), extracting respectively:

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

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

3. Decoder (Three Components)

(a) Face Transformer: Processes candidate features grouped by viewpoint. Each viewpoint group \(f_v^S\) first undergoes cross-attention with the global feature \(f^I\) (injecting global shape priors), followed by intra-group self-attention (enhancing local geometric consistency). Viewpoint positional encodings \(E_v^{fpos}\) are used to preserve spatial relationships.

(b) Classification Head: Applies MLP + softmax to the global feature \(f^I\) to predict a class probability distribution \(\mathbf{p}^{cls} \in \mathbb{R}^c\), providing class-level semantic priors for completion.

(c) Fusion Network: The geometric features \(F^S\) are first processed by a compress-expand MLP (bottleneck dimension of 4, aligned with the output dimension), then concatenated with the class probabilities \(\mathbf{p}^{cls}\). A prediction head then outputs for each candidate point:

• Offset \(o_m = \{o_x, o_y, o_z\}\): a positional correction in the local coordinate frame centered at the candidate point
• Opacity \(\sigma_m\): determines whether the point is valid (retained if \(\sigma \geq 0.5\))

The final completed 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 denotes Chamfer Distance and \(\mathcal{L}_{cls}\) is the cross-entropy classification loss.

Key Experimental Results

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

MVP Dataset (16 classes)

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

Both settings outperform prior SOTA methods including DualGenerator and PDR.

Ablation Study (PCN)

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

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

Robustness

• Noise robustness: Under varying levels of Gaussian noise, DANCE exhibits smaller performance degradation compared to SVDFormer and SeedFormer.
• Density flexibility: Trained with fixed \(R=21\), the model can directly use \(R=17\) or \(R=29\) at inference to adjust output density without retraining.

Highlights & Insights

1. Density-agnostic design: The first point cloud completion method to achieve density-agnosticism for both input and output, with output point count naturally controlled via opacity filtering.
2. Pure 3D semantic priors: The classification head learns class information directly from 3D geometric features without image supervision, making it more suitable for real-world deployment.
3. Completion of missing regions only: Original observed geometry is preserved, avoiding detail loss introduced by global regeneration.
4. Ray-based sampling strategy: The NeRF-inspired idea is elegantly transferred to point cloud completion, providing a structured distribution of candidate points.
5. Inference-time density control: Output resolution can be flexibly adjusted by modifying \(R\), offering strong practical utility.

Limitations & Future Work

1. Fixed viewpoint configuration: The current hexahedral viewpoint layout and uniform grid may not constitute an optimal sampling strategy for highly asymmetric or geometrically complex objects.
2. Computational overhead from candidate points: When \(R\) is large, \(M = V \cdot R^2\) generates a substantial number of candidate points, increasing computational cost.
3. Limited category set: The classification head relies on a predefined set of categories, raising concerns about generalization to unseen classes.
4. Evaluation on synthetic data only: Both PCN and MVP are ShapeNet-based synthetic benchmarks; the method has not been validated on real sensor scans (e.g., ScanNet, KITTI).
5. The authors propose adaptive viewpoint sampling as a future direction, dynamically adjusting sampling positions and viewpoint count based on input geometry.

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	Not supported	Not supported	Adjustable \(R\) at inference

DANCE belongs to the same "complete missing regions only" category as PoinTr, but achieves density flexibility through the opacity mechanism and, with the addition of semantic priors, reduces CD-Avg on PCN from PoinTr's 7.76 to 6.46.

The following broader insights are noteworthy:

• Cross-domain transfer from NeRF to point cloud completion: The combination of ray-based sampling and opacity prediction can be generalized to other 3D generation tasks (e.g., scene completion, point cloud upsampling).
• Lightweight semantic guidance: A simple classification head yields significant improvements in completion quality, suggesting that semantic priors can be incorporated at low cost in other 3D tasks as well.
• Density-controllable inference: The opacity filtering mechanism can be adopted in scenarios requiring flexible output resolution control (e.g., level-of-detail generation).

Rating

• Novelty: ⭐⭐⭐⭐ — The density-agnostic design and NeRF-style sampling transfer are novel, though the classification head itself is relatively straightforward
• Experimental Thoroughness: ⭐⭐⭐⭐ — Comprehensive comparisons on PCN/MVP with complete ablations, but real-world data validation is lacking
• Writing Quality: ⭐⭐⭐⭐ — Motivation is clearly articulated, structure is well-organized, and figures effectively aid understanding
• Value: ⭐⭐⭐⭐ — Density-agnostic completion addresses a practically meaningful challenge with clear implications 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] Simba: Towards High-Fidelity and Geometrically-Consistent Point Cloud Completion via Transformation Diffusion
• [AAAI 2026] ASSIST-3D: Adapted Scene Synthesis for Class-Agnostic 3D Instance Segmentation
• [AAAI 2026] Class-Partitioned VQ-VAE and Latent Flow Matching for Point Cloud Scene Generation