CC-4DGS: Computational Deformation and Point-Cloud Compression for Storage-Efficient Dynamic Gaussian Splatting

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Dynamic four-dimensional (4D) Gaussian Splatting (4D GS) excels at high-quality view synthesis.
• Existing 4D GS methods suffer from large storage requirements (tens to hundreds of MB per scene).
• This is primarily due to large multi-resolution hash tables and high-dimensional Gaussian attributes.

Purpose of the Study:

• To develop a storage-efficient and scalable framework for 4D Gaussian Splatting.
• To reduce the memory footprint of 4D GS representations without compromising rendering quality.
• To enable efficient storage and real-time rendering of dynamic scenes.

Main Methods:

• Introduced a Computational Deformation Field (CDF) using deterministic dense hash encoding and neural decoders to replace large hash tables.
• Developed a Compression of Canonical Point-cloud Attributes (CCA) pipeline utilizing conditional autoencoding, quantization, and residual codebooks.
• Integrated CDF and CCA into a unified representation for dynamic scenes.

Main Results:

• Reduced deformation storage to 1-3 MB per scene.
• Achieved 3-5x reduction in point-cloud attributes with negligible quality loss.
• Reduced total scene storage to 20-30 MB while maintaining real-time rendering performance.
• Demonstrated comparable reconstruction accuracy to state-of-the-art methods like Swift4D.

Conclusions:

• CC-4DGS offers a significant improvement in storage efficiency for 4D Gaussian Splatting.
• The framework provides favorable runtime-memory trade-offs for dynamic scene representation.
• CC-4DGS enables high-quality view synthesis with substantially reduced storage demands.