GaussianImage: 1000 FPS Image Representation and Compression by 2D Gaussian Splatting (2403.08551v5)

Abstract: Implicit neural representations (INRs) recently achieved great success in image representation and compression, offering high visual quality and fast rendering speeds with 10-1000 FPS, assuming sufficient GPU resources are available. However, this requirement often hinders their use on low-end devices with limited memory. In response, we propose a groundbreaking paradigm of image representation and compression by 2D Gaussian Splatting, named GaussianImage. We first introduce 2D Gaussian to represent the image, where each Gaussian has 8 parameters including position, covariance and color. Subsequently, we unveil a novel rendering algorithm based on accumulated summation. Remarkably, our method with a minimum of 3$\times$ lower GPU memory usage and 5$\times$ faster fitting time not only rivals INRs (e.g., WIRE, I-NGP) in representation performance, but also delivers a faster rendering speed of 1500-2000 FPS regardless of parameter size. Furthermore, we integrate existing vector quantization technique to build an image codec. Experimental results demonstrate that our codec attains rate-distortion performance comparable to compression-based INRs such as COIN and COIN++, while facilitating decoding speeds of approximately 2000 FPS. Additionally, preliminary proof of concept shows that our codec surpasses COIN and COIN++ in performance when using partial bits-back coding. Code is available at https://github.com/Xinjie-Q/GaussianImage.

• The paper introduces a 2D Gaussian splatting method that models images with 8-parameter Gaussians, significantly reducing computational overhead.
• It develops an accumulated blending-based rasterization and vector quantization-based codec, achieving a fast 1000 FPS decoding speed.
• Experimental results on Kodak and DIV2K datasets demonstrate at least a 3x GPU memory reduction and competitive rate-distortion performance compared to traditional INRs.

Exploring Efficient Image Representation with 2D Gaussian Splatting in GaussianImage

Overview

Researchers in the field of computer vision and image processing have been relentlessly exploring more efficient techniques for image representation and compression. Recent advancements in Implicit Neural Representations (INRs) have showcased impressive capabilities in capturing high-fidelity details of images, albeit with the drawback of requiring sizeable computational resources. Addressing the limitations prevalent in current INRs, particularly their substantial GPU memory and computational demands, a novel paradigm—termed GaussianImage—has been proposed. This new methodology centers around the use of 2D Gaussian Splatting for image representation and compression, presenting a significant shift from traditional MLP-based or feature grid-based neural representations.

Novel Contributions

GaussianImage introduces several key innovations that enhance both the practical and theoretical aspects of image representation:

• 2D Gaussian Representation: By adopting 2D Gaussians instead of the conventional 3D, a concise representation is achieved. Each Gaussian is characterized by 8 parameters, reducing the parameter count and, consequently, the computational overhead significantly.
• Accumulated Blending-based Rasterization: A novel rasterization mechanism replacing the depth-sorted alpha blending with an accumulated summation technique. This bypasses the need for Gaussian sorting based on depth information, thus streamlining the rendering process.
• Vector Quantization-based Codec: The transition of 2D Gaussian representation into an image codec through attribute quantization-aware fine-tuning and encoding showcases superior decoding speeds (around 1000 FPS) while maintaining competitive rate-distortion performance.
• Use of Partial Bits-Back Coding: Although positioned as a preliminary proof of concept, this aspect holds the promise for further bitrate reductions, potentially setting new benchmarks for compression efficiency.

Experimental Validation

Comprehensive evaluations on standard datasets (Kodak and DIV2K) against a variety of baseline methods demonstrate the robustness and efficiency of GaussianImage. Not only does it offer a reduction in GPU memory usage by a minimum of 3x and a 5x faster fitting time, but it also delivers the fastest rendering speed observed, irrespective of parameter size. Notably, the proposed 2D Gaussian Splatting approach outperforms existing INRs in representation performance while accomplishing substantially faster training and inference speeds.

When deployed as an image codec, GaussianImage exhibits competitive rate-distortion performance against established compression-based INR methods like COIN and COIN++, further distinguished by its significantly faster decoding speed. Moreover, the preliminary incorporation of partial bits-back coding hints at the potential for even further performance enhancements in terms of compression efficiency.

Implications and Future Directions

The GaussianImage framework signifies a paradigm shift in image representation, highlighting the potential of leveraging 2D Gaussian Splatting over conventional INRs or 3D splatting approaches. Its efficiency in representation and fast decoding opens new avenues for deploying high-performance image codecs on devices with varying computational capabilities. Moving forward, the exploration could extend to further optimizing the ratio of performance to computational demands and examining the utility of GaussianImage in broader domains such as video compression and real-time streaming applications.

In conclusion, the introduction of GaussianImage marks a significant stride towards realizing efficient and practical solutions for image representation and compression. Its blend of innovative methodologies promises enhancements in computational efficiency and speed, setting the stage for future explorations in the field of generative AI and beyond.