Point Cloud Completion Network Based on Multi-Dimensional Adaptive Feature Fusion and Informative Channel Attention Mechanism

Di Tian, Jiahang Shi, Jiabo Li, Mingming GongSensors 2025 · DOI

TL;DR

This paper introduces FFPF-Net, a novel point cloud completion network that tackles the common issue of missing local details in 3D reconstructions. By combining a multi-level feature extraction module (SCMP) and an attention-based refinement module (SEP-Net), FFPF-Net significantly improves the accuracy and detail of completed point clouds, outperforming existing methods.

What Problem Does This Paper Address?

Point clouds, essential for 3D perception in applications like autonomous driving and robotics, are often incomplete due to sensor limitations, occlusions, or environmental factors. These "holes" in the data severely degrade the performance of downstream tasks. While existing methods often focus on reconstructing the overall shape, they tend to overlook crucial local structural details, leading to incomplete or inaccurate reconstructions. This lack of fine-grained detail can be critical for tasks requiring precise object boundary recognition or intricate surface feature recovery.

Methods

The core of this work lies in two novel modules integrated into their proposed Feature Fusion Point Fractal Network (FFPF-Net):

  • Set Combination Multi-Layer Perceptron (SCMP): This module addresses the loss of local detail by simultaneously extracting both local and global features. It employs a multi-level, multi-resolution farthest point sampling strategy to select key points, followed by local grouping and feature extraction using Composite Multi-Layer Perceptrons (CMLP). This approach aims to preserve structural clues and provide rich local semantics.
  • Squeeze Excitation Pooling Network (SEP-Net): To enhance feature representation and refine details, SEP-Net acts as an informative channel attention mechanism. It integrates "Squeeze" (global average pooling) and "Excitation" (fully connected layers with sigmoid activation) operations to adaptively identify and enhance critical channel features, effectively recalibrating feature maps without significant computational overhead.

These modules are combined in an end-to-end FFPF-Net architecture that fuses multi-dimensional features for enhanced representation and progressively refines missing regions.

Datasets

The authors evaluated their method on two standard point cloud datasets:

  • ShapeNet-Part: A large-scale, multi-category 3D model database suitable for detailed shape understanding.
  • MVP Dataset: A multi-view partial point cloud dataset known for its large scale and diverse object categories, offering rich data for completion and registration tasks.

Key Insights and Ideas

  • Local Details Matter: The paper strongly emphasizes that simply reconstructing global shapes is insufficient. Fine-grained local details are critical for accurate 3D perception and real-world applications.
  • Multi-Level Feature Extraction is Key: The SCMP module demonstrates that a hierarchical approach to feature extraction, considering different scales and resolutions, is vital for capturing both global context and local intricacies.
  • Attention for Refinement: SEP-Net highlights the power of channel attention mechanisms in selectively amplifying important features and suppressing noise, leading to more robust and detailed reconstructions.
  • Fusion for Completeness: The FFPF-Net architecture's strength lies in its ability to fuse features from different dimensions and scales, enabling a more comprehensive and accurate completion process.

Key Results

The FFPF-Net demonstrated significant improvements over existing methods like L-GAN and PCN:

  • Average Prediction Error Improvement: The proposed method achieved average prediction error improvements of 1.3 and 1.4 compared to L-GAN and PCN, respectively.
  • Average Completion Errors: On ShapeNet-Part and MVP datasets, the average completion errors were 0.783 and 0.824, respectively, indicating superior fine-detail reconstruction.
  • Robustness: Experiments showed strong robustness in handling point clouds with varying degrees of missing data (25%, 50%, 75% loss).
  • Efficiency: The FFPF-Net model has a size of 2.8 M parameters and an inference time of 6.5 ms, making it competitive and adaptable for resource-constrained environments.
  • Qualitative Results: Visualizations showed FFPF-Net effectively restoring missing parts (e.g., aircraft tail, hat corners, chair backrests, guitar tops) with greater accuracy and detail compared to baseline methods.

Limitations and Future Work

The authors acknowledge several limitations and propose future research directions:

  • Predefined Key Point Selection: The reliance on predefined key point selection in SCMP could be a limitation for severely missing or abnormally irregular shapes. Future work aims to explore learnable deformable attention modules for dynamic region importance determination.
  • Model Complexity and Efficiency: While FFPF-Net offers improved performance, its complexity is higher than simpler models. Future research will focus on channel pruning and dynamic inference paths to create a dual-mode system for adaptive precision and speed.
  • Real-World Robustness: Experiments were confined to benchmark datasets. Future work will involve building real-time data collection platforms and developing anti-interference strategies (e.g., noise simulation) to validate performance in complex physical environments.
  • Inference Speed: While efficient, further optimizations for inference speed are needed for high-real-time applications like autonomous driving.

Summary updated

Abstract

With the continuous advancement of 3D perception technology, point cloud data has found increasingly widespread application. However, the presence of holes in point cloud data caused by device limitations and environmental interference severely restricts algorithmic performance, making point cloud completion a research topic of high interest. This study observes that most existing mainstream point cloud completion methods primarily focus on capturing global features, while often underrepresenting local structural details. Moreover, the generation process of complete point clouds lacks effective control over fine-grained features, leading to insufficient detail in the completed outputs and reduced data integrity. To address these issues, we propose a Set Combination Multi-Layer Perceptron (SCMP) module that enables the simultaneous extraction of both local and global features, thereby reducing the loss of local detail information. In addition, we introduce the Squeeze Excitation Pooling Network (SEP-Net) module, an informative channel attention mechanism capable of adaptively identifying and enhancing critical channel features, thus improving the overall feature representation capability. Based on these modules, we further design a novel Feature Fusion Point Fractal Network (FFPF-Net), which fuses multi-dimensional point cloud features to enhance representation capacity and progressively refines the missing regions to generate a more complete point cloud. Extensive experiments conducted on the ShapeNet-Part and MVP datasets compared to L-GAN and PCN showed average prediction error improvements of 1.3 and 1.4, respectively. The average completion errors on the ShapeNet-Part and MVP datasets are 0.783 and 0.824, highlighting the improved fine-detail reconstruction capability of our network. These results indicate that the proposed method effectively enhances point cloud completion performance and can further promote the practical application of point cloud data in various real-world scenarios.
BibTeX
@inproceedings{tian2025point,
  author = {Di Tian and Jiahang Shi and Jiabo Li and Mingming Gong},
  title = {{Point Cloud Completion Network Based on Multi-Dimensional Adaptive Feature Fusion and Informative Channel Attention Mechanism}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25196173},
}
Figure