Reflective Gaussian Splatting
Authors: Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, Li Zhang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on standard datasets demonstrate that Ref Gaussian surpasses existing approaches in terms of quantitative metrics, visual quality, and compute efficiency. Further, we show that our method serves as a unified solution for both reflective and non-reflective scenes, going beyond the previous alternatives focusing on only reflective scenes. Also, we illustrate that Ref-Gaussian supports more applications such as relighting and editing. |
| Researcher Affiliation | Academia | 1School of Data Science, Fudan University 2University of Surrey Co-last authorship BLi Zhang (EMAIL) is the corresponding author. |
| Pseudocode | No | The paper describes its methodology using textual explanations, mathematical equations, and diagrams. It does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | https://fudan-zvg.github.io/ref-gaussian. The provided URL is a project page (GitHub Pages), not a direct link to a source-code repository, and the paper does not contain an explicit statement about releasing the code for the described methodology. |
| Open Datasets | Yes | Datasets We select two synthetic datasets Shiny Blender (Verbin et al., 2022) and Glossy Synthetic (Liu et al., 2023) for novel view synthesis of reflective objects, and Ref-Real dataset (Verbin et al., 2022) for real-world open scenes. Competitors We compare several representative models in the field of reflective 3D reconstruction, including Ref-Ne RF (Verbin et al., 2022), ENVIDR (Liang et al., 2023), 3DGS (Kerbl et al., 2023), Gaussian Shader (Jiang et al., 2024), Relightable Gaussian (R3DG) (Gao et al., 2023), and 3DGSDR (Ye et al., 2024). We also compared with more Ne RF-based models like Ne RO (Liu et al., 2023), Neu S (Wang et al., 2021) and NDE (Wu et al., 2024), see Tables 5 and 6 in appendix. Evaluation metrics We use three standard metrics: PSNR, SSIM (Wang et al., 2004), and LPIPS (Zhang et al., 2018). |
| Dataset Splits | No | The paper mentions using synthetic and real-world datasets for novel view synthesis and comparing against other models, evaluating on 'synthesized test views'. However, it does not provide specific details on how these datasets were split into training, validation, or test sets (e.g., percentages, sample counts, or explicit reference to standard splits used by the authors). |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA A6000 GPU. |
| Software Dependencies | No | The paper does not explicitly state specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9, CUDA 11.1) that would be needed to reproduce the experiments. |
| Experiment Setup | Yes | Our training procedure consists of two stages. We utilize a per-Gaussian rendering for 18, 000 steps as an initial stage, followed by a deferred rendering stage training for about 40, 000 steps. We reset all color and material attributes before the second stage, retaining only the the geometry of Gaussians. The learning rates of the trainable material attributes (metallic, roughness, and albedo) are all set to 0.005, while the learning rate of the environment map is set to 0.01. The set of other basic trainable attributes for Gaussians like position and covariance is consistent with 2DGS (Huang et al., 2024). Then, our material-aware normal propagation is conducted to those Gaussians with metallic no less than 0.02 and roughness no more than 0.1. Also, we adopt the metallic initial value from 3DGS-DR (Ye et al., 2024) and set the initial roughness value to be 0.1. During the implementation of physically based rendering, we discovered that spherical harmonics have a better fitting capability than the integrated diffuse lighting and we thereby use spherical harmonics to substitute it. For the loss function, λn is set to 0.05 and λsmooth is 1.0. Additionally, we periodically extract object s surface mesh at 3000 step intervals. |