Compression of 3D Gaussian Splatting with Optimized Feature Planes and Standard Video Codecs

Soonbin Lee, Fangwen Shu, Yago Sanchez de la Fuente, Thomas Schierl, Cornelius Hellge
Fraunhofer Heinrich Hertz Institute, HHI
arXiv Code (Coming Soon)
MY ALT TEXT

TL;DR

We propose a compact 3D Gaussian Splatting (3DGS) scene representation based on a grid-based feature plane model that incorporates (1) frequency-domain entropy modeling and (2) channel importance bit allocation with a progressive training scheme. The feature plane is compressed using a standard video codec (e.g., HEVC), enabling high compression rates with low complexity.

MY ALT TEXT

Visualization of the first channel in the XY plane (a) with L1 loss and (b) with proposed DCT entropy loss.
Our experimental results show that DCT entropy loss improves rate-distortion performance with standard video codecs.

Key Insight

  1. Feature Plane Representation: Rather than storing all primitive attributes explicitly in 3DGS, we use feature planes to predict them. We employ progressive training that limits backpropagation at each channel level, which enhances training stability.
  2. Entropy Loss for Video Codec: Standard video codec is non-differentiable, so we directly optimize the block-wise DCT coefficient of feature plane by entropy modeling. By introducing this loss, the bitstream size of the feature plane is significantly reduced without requiring any codec software modifications.
  3. Channel Importance-based Bit Allocation: Based on progressive training, we calculate importance with different channel levels and incorporate them into the optimization loss. The proposed bit allocation method improves compression efficiency.

MY ALT TEXT

Quantitative results for the compression of 3DGS. We conducted experiments using the Mip-NeRF360, DeepBlending, and Tank & Temples datasets. The results are highlighted in red, orange and yellow, respectively.

Visual Comparisons

Ours, 15.2MB
Compact3DGS (CVPR24), 43.2MB
Ours, 7.7MB
Compact3DGS (CVPR24), 24.6MB
Ours, 7.9MB
CompGS (MM24), 8.9MB
Ours, 8.8MB
CompGS (MM24), 10.1MB

BibTeX

@misc{lee2025compression3dgaussiansplatting,
      title={Compression of 3D Gaussian Splatting with Optimized Feature Planes and Standard Video Codecs}, 
      author={Soonbin Lee and Fangwen Shu and Yago Sanchez and Thomas Schierl and Cornelius Hellge},
      year={2025},
      eprint={2501.03399},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.03399}, 
}