ASH: A Modern Framework for Parallel Spatial Hashing in 3D Perception.

Area of Science:

• Computer Vision
• High-Performance Computing
• GPU Computing

Background:

• Spatial hashing is crucial for many 3D computer vision tasks.
• Existing GPU hash map implementations often lack performance, flexibility, or ease of use.
• Implementing spatially varying operations requires efficient data handling.

Purpose of the Study:

• Introduce ASH, a high-performance parallel spatial hashing framework for GPUs.
• Provide a versatile tensor interface to simplify the implementation of complex 3D operations.
• Enhance performance and reduce code complexity compared to existing solutions.

Main Methods:

• Decoupled internal hashing data structures from key-value data.
• Utilized an index heap to bridge low-level pointer-first structures to high-level tensor interfaces.
• Adapted integer-only hash map implementations for multi-dimensional keys.

Main Results:

• ASH demonstrates significant performance gains over state-of-the-art hash maps on synthetic data.
• Achieved higher performance with fewer lines of code (LoC) on large-scale 3D perception tasks.
• Successfully applied ASH to point cloud voxelization, volumetric scene reconstruction, and geometry/appearance refinement.

Conclusions:

• ASH offers a high-performance, flexible, and user-friendly solution for parallel spatial hashing on GPUs.
• The framework's design enables seamless integration with modern libraries like PyTorch.
• ASH facilitates advanced 3D computer vision applications and is open-sourced in Open3D.