Grid4D: 4D Decomposed Hash Encoding for High-Fidelity Dynamic Gaussian Splatting
Authors: Jiawei Xu, Zexin Fan, Jian Yang, Jin Xie
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we introduce our experiments conducted on a single RTX 3090 GPU. We build our code mainly on Py Torch [30], while we implement our 4D decomposed hash encoder with CUDA/C++. More experimental results and analysis can be found in the supplementary. |
| Researcher Affiliation | Academia | Jiawei Xu1 , Zexin Fan1 , Jian Yang1 , Jin Xie23 1PCA Lab, VCIP, College of Computer Science, Nankai University 2State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China 3School of Intelligence Science and Technology, Nanjing University, Suzhou, China {jiaweixu, zexin_fan}@mail.nankai.edu.cn csjxie@nju.edu.cn csjyang@nankai.edu.cn |
| Pseudocode | No | The paper does not contain a pseudocode block or an algorithm section explicitly labeled as such. |
| Open Source Code | No | Project page: https://jiaweixu8.github.io/Grid4D-web/. We will release all our codes on the Git Hub platform, and provide the URL in the camera-ready version. We will offer the instructions to reproduce the results in detail. |
| Open Datasets | Yes | We evaluate Grid4D on two popular datasets. D-Ne RF [31] dataset is a public monocular synthetic dataset that provides accurate and time-varying camera poses. Hyper Ne RF [28] dataset is a public real-world dataset captured by one or two moving cameras. Neu3D [16] dataset is a public dataset captured by multiple cameras with fixed poses. |
| Dataset Splits | No | Quantitative comparison on the validation rig part (Rig) and the interpolation part (Interpolation) of the real-world Hyper Ne RF [28] dataset. The paper mentions evaluating on a 'validation rig part' for one dataset, but does not provide explicit train/validation/test splits (e.g., percentages or counts) for its training procedure across datasets. |
| Hardware Specification | Yes | In this section, we introduce our experiments conducted on a single RTX 3090 GPU. |
| Software Dependencies | No | We build our code mainly on Py Torch [30], while we implement our 4D decomposed hash encoder with CUDA/C++. |
| Experiment Setup | Yes | For all datasets, we configure the resolution of the spatial grid hash encoder to span from 16 to 2048 across 16 levels. Meanwhile, the max level number L of temporal grid hash encoders remains consistent at 32. We set λc and λr to 0.2 and 0.5 for common scenes and follow a similar learning rate schedule as Deform GS [50, 13]. |