# EAGLES: Efficient Accelerated 3D Gaussians with Lightweight EncodingS

**Authors:** Sharath Girish, Kamal Gupta, Abhinav Shrivastava  
**arXiv:** [2312.04564](https://arxiv.org/abs/2312.04564)  
**Date:** 2023-12-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.04564v1/

## Key points

- 3D Gaussian splatting (3D-GS) is a popular technique for novel-view scene synthesis
- It addresses the challenges of lengthy training times and slow rendering speeds associated with Neural Radiance Fields (NeRFs)
- 3D-GS achieves real-time rendering and accelerated training by rasterizing 3D Gaussians in a differentiable manner
- However, it requires significant memory resources due to the need for millions of Gaussians in the point cloud representation of each scene
- EAGLES (Efficient Accelerated 3D Gaussians with Lightweight EncodingS) is proposed as a solution to this limitation
- EAGLES utilizes quantized embeddings to reduce memory storage requirements and introduces a coarse-to-fine training strategy for faster optimization of Gaussian point clouds
- Results show impressive improvements in memory efficiency and computational speed while maintaining high reconstruction quality
- EAGLES consistently outperforms existing methods in terms of PSNR and SSIM
- It achieves better storage efficiency with reduced memory consumption ranging from 9MB to 746MB across different scenarios
- EAGLES also demonstrates improved performance in terms of FPS during rendering and reduced training time compared to other methods
- Overall, EAGLES presents a highly efficient and effective approach for real-time rendering of high-resolution scenes

## Summary

Recently, 3D Gaussian splatting (3D-GS) has emerged as a popular technique for novel-view scene synthesis. It addresses the challenges of lengthy training times and slow rendering speeds associated with Neural Radiance Fields (NeRFs). By rapidly rasterizing 3D Gaussians in a differentiable manner, 3D-GS achieves real-time rendering and accelerated training. However, one major drawback is the significant memory resources required for both training and storage due to the need for millions of Gaussians in the point cloud representation of each scene. To overcome this limitation, the authors propose a novel technique called EAGLES (Efficient Accelerated 3D Gaussians with Lightweight EncodingS). The key idea behind EAGLES is to utilize quantized embeddings to significantly reduce memory storage requirements while introducing a coarse-to-fine training strategy that enables faster and more stable optimization of the Gaussian point clouds. The results of their approach demonstrate impressive improvements in terms of memory efficiency and computational speed; they achieve a reduction in memory usage by more than an order of magnitude while maintaining high reconstruction quality. The effectiveness of their approach is validated on various datasets and scenes, showcasing its ability to preserve visual quality while consuming significantly less memory and achieving faster training/inference speeds. In comparison to existing methods such as Plenoxels, INGP, M-NeRF360, and 3D-GS*, EAGLES consistently outperforms them in terms of PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure). Moreover, EAGLES achieves better storage efficiency with reduced memory consumption ranging from 9MB to 746MB across different scenarios. The proposed technique also demonstrates improved performance in terms of FPS (Frames Per Second) during rendering and reduced training time compared to other methods; for example, EAGLES achieves an FPS ranging from 7 to 283 depending on the dataset and scene while reducing training time from hours to minutes. Overall, EAGLES presents a highly efficient and effective approach for real-time rendering of high resolution scenes. Its quantized embeddings and coarse-to-fine training strategy contribute to significant memory savings and faster training/inference speeds without compromising reconstruction quality. The availability of the project page and code further facilitates the adoption and implementation of this technique in various applications.