GI-GS: Global Illumination Decomposition on Gaussian Splatting for Inverse Rendering

Authors: Hongze CHEN, Zehong Lin, Jun Zhang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Quantitative and qualitative results show that our GI-GS outperforms existing baselines in both rendering quality and efficiency. Project page: https://stopaimme.github.io/GI-GS-site/. ... We perform experiments on the Tenso IR synthetic dataset (Jin et al., 2023) and the Mip-Ne RF 360 (Mildenhall et al., 2021) dataset. For both datasets, we evaluate the quality of novel view synthesis to compare the overall performance. ... To quantify the performance, we employ three key metrics: PSNR, SSIM, and LPIPS. ... 5.3 ABLATION STUDY
Researcher Affiliation Academia Hongze Chen, Zehong Lin , Jun Zhang The Hong Kong University of Science and Technology EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes methods and formulas (e.g., equations for rendering and ray tracing) but does not include any explicitly labeled pseudocode or algorithm blocks with structured steps.
Open Source Code Yes Project page: https://stopaimme.github.io/GI-GS-site/.
Open Datasets Yes We perform experiments on the Tenso IR synthetic dataset (Jin et al., 2023) and the Mip-Ne RF 360 (Mildenhall et al., 2021) dataset.
Dataset Splits No The paper mentions using the Tenso IR synthetic dataset and the Mip-Ne RF 360 dataset for experiments, evaluating novel view synthesis. However, it does not explicitly provide specific details on how these datasets were split into training, validation, or test sets within the provided text.
Hardware Specification Yes All training is conducted on a single NVIDIA A5000 GPU.
Software Dependencies No The paper does not explicitly state specific software dependencies with version numbers (e.g., Python version, specific library versions like PyTorch 1.x or CUDA 11.x).
Experiment Setup Yes The training iterations are set to 30K, with the same learning rate as the vanilla 3DGS. ... The optimization of the materials and the lighting takes 10,000 and 5,000 iterations on the Mip-Ne RF 360 and Tenso IR datasets, respectively. For the BRDF attributes, we adopt the continuous learning rate decay strategy in Plenoxels(Fridovich-Keil et al., 2022), where the initial learning rate is set to 0.05 and decays to a final learning rates of 0.005. ... Following GS-IR, we set λn T V , λM, and λE to 5.0, 1.0, and 0.01, respectively.