Lossless Coding of Multi-Resolution Hash Tables for Instant-NGP Representations of 3D Scenes

This work presents specialized lossless coding techniques for Instant-NGP's hash tables, achieving 5.27% bit rate reduction over the MPEG Neural Network Coding (NNC) standard. In Instant-NGP, small MLPs combined with hash tables are trained to render 3D scenes. These hash tables consist of learnable feature vectors, organized hierarchically across levels of increasing resolution. Only the hash tables were quantized and compressed, since they account for over 99% of the model parameters.