R$^2$-Gaussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction

Authors: Ruyi Zha, Tao Jun Lin, Yuanhao Cai, Jiwen Cao, Yanhao Zhang, Hongdong Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on synthetic and real-world datasets demonstrate that our method outperforms state-of-the-art approaches in accuracy and efficiency.
Researcher Affiliation Academia 1The Australian National University 2Johns Hopkins University 3Robotics Institute, University of Technology Sydney
Pseudocode No The paper describes algorithms in text but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes Code and models are available on the project page https://github.com/Ruyi-Zha/r2_gaussian.
Open Datasets Yes For the synthetic dataset, we collect 15 real CT volumes... For real-world experiments, we use three cases from the FIPS dataset [56]... Refer to Appendix B for more details of datasets. (Appendix B: LIDC-IDRI [4] and Pancreas-CT [47], X-Plant [58], Sci Vis [26], FIPS [56])
Dataset Splits No The paper discusses training and testing but does not explicitly provide training/test/validation dataset splits or refer to a validation set.
Hardware Specification Yes All methods run on a single RTX3090 GPU.
Software Dependencies No Our R2-Gaussian is implemented in Py Torch [44] and CUDA [50]... (no version numbers provided for reproducibility)
Experiment Setup Yes Learning rates for position, density, scale, and rotation are initially set as 0.0002, 0.01, 0.005, and 0.001, respectively, and exponentially to 0.1 of their initial values. Loss weights are λssim = 0.25 and λtv = 0.05. We initialize M = 50k Gaussians with a density threshold τ = 0.05 and scaling term k = 0.15. The TV volume size is D = 32. Adaptive control runs from 500 to 15k iterations with a gradient threshold of 0.00005.