Dynamic-focus transformer for point cloud segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Transformer models excel at 3D point cloud segmentation by capturing long-range dependencies.
• Existing methods face computational redundancy and overfitting due to global or fixed-window self-attention.

Purpose of the Study:

• To develop a novel, efficient, and accurate transformer architecture for 3D point cloud segmentation.
• To address the limitations of computational redundancy and overfitting in current transformer-based approaches.

Main Methods:

• Proposed the Dynamic-Focus Transformer with a data-dependent adaptive attention mechanism.
• Utilized learned soft point masks to selectively sparsify keys and values, focusing on critical regions.
• Integrated the method into a U-Net-style encoder-decoder for an efficient balance of capability and cost.

Main Results:

• Achieved state-of-the-art performance on S3DIS and ScanNetv2 benchmarks.
• Demonstrated notably improved efficiency compared to existing methods.
• Validated the effectiveness for large-scale point cloud understanding.

Conclusions:

• The Dynamic-Focus Transformer offers a flexible, input-adaptive receptive field without high memory overhead.
• The proposed architecture provides a highly efficient balance between modeling capability and computational cost for 3D point cloud segmentation.