LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS

Authors: Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, Zhangyang "Atlas" Wang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comparisons using the scene-scale view synthesis dataset provided by Mip-Ne RF360 [64], which comprises nine real-world large-scale scenes... We report metrics including the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity as measured by LPIPS [66]. ...We performed ablation studies on each component of our method to evaluate their individual impacts.
Researcher Affiliation Academia Zhiwen Fan1 , Kevin Wang1 , Kairun Wen2, Zehao Zhu1, Dejia Xu1, Zhangyang Wang1 1The University of Texas at Austin 2XMU
Pseudocode Yes Algorithm 1 The overall pipeline of Light Gaussian
Open Source Code No We will release our code after our paper gets accepted.
Open Datasets Yes We conduct comparisons using the scene-scale view synthesis dataset provided by Mip-Ne RF360 [64], which comprises nine real-world large-scale scenes... In addition, we utilize the Tanks and Temples dataset [65]... The synthetic Blender dataset [5] includes eight photo-realistic synthetic objects with ground-truth controlled camera poses and rendered viewpoints (100 for training and 200 for testing).
Dataset Splits No The paper does not specify validation splits (percentages or counts) or cite predefined splits for any dataset.
Hardware Specification Yes All performance evaluations are conducted on an A6000 GPU.
Software Dependencies No Our framework is implemented in Py Torch and integrates the differentiable Gaussian rasterization technique from 3D-GS [16].
Experiment Setup Yes In the Global Significance Calculation phase, we assign a power value of 0.1 in Eq. 4 and proceed to fine-tune the model for 5,000 steps during the Gaussian Co-adaptation process. For SH distillation, we downscale the 3-degree SHs to 2-degree, thereby reducing 21 elements for each Gaussian. This is further optimized by setting σ to 0.1 in the pseudo view synthesis stage. In the Gaussian VQ step, the codebook size is configured to 8192, selecting SHs with the least 60% significance score for the vector quantization (VQ ratio), to balance the trade-off between compression efficiency and fidelity.