H-CNN: Spatial Hashing Based CNN for 3D Shape Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• 3D shape analysis using deep learning often requires significant memory and computational resources.
• Existing methods struggle with memory efficiency, especially at high resolutions.
• Efficient parallelization of convolutional neural network (CNN) operations on 3D data is challenging.

Purpose of the Study:

• To develop a novel spatial hashing data structure for efficient 3D shape analysis using CNNs.
• To reduce the memory footprint and accelerate CNN training for 3D models.
• To enable high-resolution 3D shape analysis with improved computational efficiency.

Main Methods:

• A hierarchical hash table data structure is built for 3D models, leveraging sparse occupancy of shape boundaries.
• Perfect spatial hashing is employed, ensuring no collisions and minimal data structure size.
• Two GPU algorithms, hash2col and col2hash, are designed for efficient parallelization of CNN operations (convolution, pooling).

Main Results:

• The proposed data structure significantly reduces memory consumption during CNN training, especially at high resolutions (e.g., 256^3).
• CNN operations run faster due to more compact packing of input geometry features.
• Experimental results show comparable or superior benchmark performance to state-of-the-art methods with reduced memory usage (one-third at high resolutions).

Conclusions:

• The novel spatial hashing data structure offers a more memory-efficient and faster approach for 3D shape analysis with CNNs.
• This method enables high-resolution 3D analysis that was previously computationally prohibitive.
• The approach presents a significant advancement in the field of geometric deep learning and 3D computer vision.