.png){kind=icon}
.png){kind=icon}
Return
Collaborative point cloud geometry compression for both human vision and machine vision
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-025-01847-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper addresses the pressing need for efficient compression techniques for 3D point cloud data, which is crucial for both human-computer interaction and machine vision tasks. While existing methods often prioritize human perception, they fail to meet the demands of machine-driven applications, leading to data redundancy. We introduce a collaborative point cloud geometry compression approach that optimally balances human and machine vision tasks. Leveraging global and local features, our method minimizes redundancy across various tasks, achieved through a dual-branch architecture for feature extraction and entropy encoding. Another dual-branch structure facilitates high-fidelity point cloud reconstruction and machine vision tasks during decoding, with a task-friendly entropy engine enhancing coding efficiency. Our contributions include a comprehensive framework for human-machine collaborative compression, a novel dual-branch feature extraction encoder, and a coordinate reconstruction module for decoding and fusion. Extensive experiments validate the superiority of our method, promising significant reductions in bit rate while maintaining reconstruction quality and machine vision performance.
Keywords:
3D point cloud
Geometry compression
Human vision
Machine vision