Implicit Neural Representations for 3D Gaussian Compression

3D Gaussian Splatting (3DGS) distinguishes itself as cutting-edge technique for real-time novel view synthesis and is widely applied in virtual reality and augmented reality. However, 3DGS requires a large number of Gaussian primitives to represent a scene, making it difficult to be transmitted or stored. Existing methods commonly focus on directly compressing the explicit representations of 3DGS, overlooking the underlying relationship between the Gaussian primitives. To remedy this, we proposed to encode Gaussian primitives into implicit neural representation for more compact 3DGS compression. Specifically, we first build a voxel-based encoder to capture the relationship between adjacent Gaussian primitives. Then, arithmetic coding is conducted on the resultant latent representations. Next, a decoder with a point-based upsampling branch and a parallel voxel-based upsampling branch is employed to reconstruct the 3DGS. These two branches incorporate with each other to produce the final results. Experimental results demonstrate that our method can significantly reduce the storage cost as compared to 3DGS (over 80×) while maintaining competitive quality.