POTR: Post-Training 3DGS Compression

Creating 3D scenes and viewing them in real time are essential technologies in computer graphics and virtual reality. Recently, 3D Gaussian Splatting (3DGS) has emerged as the most promising contender to Neural Radiance Fields (NeRF) in 3D scene reconstruction and real-time novel view synthesis. 3DGS outperforms NeRF in training and inference speed, but falls short in storage requirements, with typical unbounded scenes requiring 250 MB to 1.5 GB of space. To remedy this downside, we present POTR, a post-training codec for 3DGS scenes that uses two novel compression techniques to achieve compression ratios around 100x while minorly affecting visual acuity. First, a modified 3DGS rasterizer accurately and efficiently calculates a splat's contribution and the change in PSNR upon its removal. Both metrics are subsequently used to remove over 80% of splats, while introducing only minor visual artifacts and without altering any other attributes. Splat removal also significantly boosts inference speeds, with scenes more than doubling in framerate. Second, a novel spherical harmonics energy compaction method is employed to lower the AC lighting coefficients' entropy substantially. Using a heavily modified version of ridge regression, up to 95% of AC lighting coefficients are set to zero while simultaneously reducing their L2 norm. Additionally, this energy compaction method can be used for a wide range of post-processing spherical harmonics operations, such as increasing or decreasing the degree of spherical harmonics after training or removing hallucinated colors.