3DGS.zip: A survey on 3D Gaussian Splatting Compression Methods

Milena T. Bagdasarian,Paul Knoll,Florian Barthel,Anna Hilsmann,Peter Eisert,Wieland Morgenstern
2024-09-03
Abstract:We present a work-in-progress survey on 3D Gaussian Splatting compression methods, focusing on their statistical performance across various benchmarks. This survey aims to facilitate comparability by summarizing key statistics of different compression approaches in a tabulated format. The datasets evaluated include TanksAndTemples, MipNeRF360, DeepBlending, and SyntheticNeRF. For each method, we report the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), and the resultant size in megabytes (MB), as provided by the respective authors. This is an ongoing, open project, and we invite contributions from the research community as GitHub issues or pull requests. Please visit <a class="link-external link-http" href="http://w-m.github.io/3dgs-compression-survey/" rel="external noopener nofollow">this http URL</a> for more information and a sortable version of the table.
Computer Vision and Pattern Recognition
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the challenges faced by 3D Gaussian Splatting (3DGS) during the compression process, namely how to maintain visual quality and real - time rendering speed while optimizing memory usage. Specifically, the paper focuses on the statistical performance of different compression methods on multiple benchmark datasets and promotes the comparability of these methods by summarizing key statistical data. ### Main Objectives and Contributions of the Paper 1. **Optimizing Memory Usage**: 3D Gaussian Splatting models usually contain a large number of Gaussian distributions, which lead to large storage requirements. The paper explores how to reduce the memory footprint of these models without affecting their rendering effects. 2. **Maintaining Visual Quality**: During the compression process, it is crucial to ensure that the visual quality of the model does not decline significantly. To this end, the paper evaluates the performance of different compression methods on metrics such as Peak Signal - to - Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). 3. **Achieving Real - Time Rendering**: In addition to memory optimization, the paper also focuses on whether the compression method can support real - time rendering, which is very important for practical applications. 4. **Transparency and Reproducibility**: To ensure the transparency of the research and the reproducibility of the results, the paper provides a detailed table listing the quantitative results of different compression methods on multiple datasets. ### Datasets and Evaluation Metrics - **Datasets**: - TanksAndTemples [9] - MipNeRF360 [1] - DeepBlending [6] - SyntheticNeRF [14] - **Evaluation Metrics**: - Peak Signal - to - Noise Ratio (PSNR) - Structural Similarity Index (SSIM) - Learned Perceptual Image Patch Similarity (LPIPS) - Compressed File Size (Size in MB) ### Research Methods The paper achieves the above objectives in the following ways: - **Redundant Gaussian Primitives Removal**: Introduce a scale - and - resolution - based redundant primitives removal method to reduce unnecessary Gaussian distributions. - **Adaptive Adjustment of Spherical Harmonic Coefficients**: Dynamically adjust the spherical harmonic (SH) bandwidth according to the minimization of view - dependent effects, thereby reducing storage requirements. - **Quantization and Encoding**: Use K - means clustering to generate a codebook and perform 16 - bit half - floating - point quantization on the remaining uncompressed floating - point values. - **End - to - End Rate - Distortion Optimization**: Achieve flexible continuous rate control through dynamic pruning and Entropy - Constrained Vector Quantization (ECVQ). - **Structured 3D Gaussian Distributions**: Utilize scene structure to guide local 3D Gaussian distributions to adapt to different viewing angles and distances. ### Summary Through a systematic investigation of multiple 3D Gaussian Splatting compression methods, this paper aims to provide a comprehensive reference for researchers and developers, helping them choose the compression technology most suitable for specific application scenarios. In addition, the paper also emphasizes the importance of transparency and reproducibility and encourages the community to contribute more research results.