MrNeRF (@janusch_patas)
2025-10-07 | โค๏ธ 225 | ๐ 13
Optimized Minimal 4D Gaussian Splatting
Contributions: โข We propose a novel multi-stage framework that progressively reduces the number of Gaussians through Gaussian Sampling, Pruning, and Merging, while maintaining reconstruction quality.
โข We generalize Sub-Vector Quantization (SVQ) for 4D representations alongside implicit appearance compression, enabling highly compact yet high-fidelity models.
โข To the best of our knowledge, we achieve state-of-the-art performance within a 3 MB memory budget, with negligible visual quality loss compared to the baseline model, condensing the model size from gigabytes to a few megabytes.
๋ฏธ๋์ด
