Point Cloud Completion Via Skeleton-Detail Transformer.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Robotics

Background:

• Point cloud shape completion is crucial for 3D vision and robotics.
• Existing methods struggle to balance global shape generation with local detail preservation.
• Convolutional approaches often overlook long-distance correlations between shape skeletons and details.

Purpose of the Study:

• To develop a novel coarse-to-fine completion framework for point clouds.
• To effectively integrate neighboring and long-distance region cues for enhanced shape restoration.
• To address the limitations of current methods in capturing global-local feature correlations.

Main Methods:

• Proposed a coarse-to-fine completion framework utilizing a Skeleton-Detail Transformer.
• Employed cross-attention and self-attention layers to explore correlations between local patterns and global shapes.
• Introduced a selective attention mechanism to optimize memory usage during the attention process.

Main Results:

• The Skeleton-Detail Transformer effectively enhances the overall shape skeleton by leveraging both local and global information.
• The selective attention mechanism reduces memory footprint without compromising performance.
• Extensive experiments on ShapeNet and real-scanned datasets show superior performance compared to state-of-the-art methods.

Conclusions:

• The proposed framework significantly improves point cloud shape completion by capturing long-distance correlations.
• The Skeleton-Detail Transformer offers a powerful approach for integrating diverse geometric cues.
• This work advances the state-of-the-art in 3D shape completion for various applications.