ODGS: 3D Scene Reconstruction from Omnidirectional Images with 3D Gaussian Splattings
Authors: Suyoung Lee, Jaeyoung Chung, Jaeyoo Huh, Kyoung Mu Lee
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive experiments highlight the superiority of ODGS by delivering the best reconstruction and perceptual quality across various datasets. Additionally, results on roaming datasets demonstrate that ODGS effectively restores fine details, even when reconstructing large 3D scenes. |
| Researcher Affiliation | Academia | Suyoung Lee 1 Jaeyoung Chung 1 Jaeyoo Huh 2 Kyoung Mu Lee 1,2 1Dept. of ECE & ASRI, 2IPAI, Seoul National University, Seoul, Korea |
| Pseudocode | No | The paper describes algorithms in descriptive text and mathematical equations but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available on our project page.1 1https://github.com/esw0116/ODGS |
| Open Datasets | Yes | We note that all datasets have CC-BY-4.0 licenses. Although some datasets provide camera poses and dense point clouds, we run the structure-from-motion, specifically Open MVG [37], on all the datasets and use obtained poses and point clouds for our experiment. |
| Dataset Splits | Yes | In our experiment, we resize them to half resolution 1920 960, and we use 20% of images for the test. Since these datasets do not split the train and test images, we conducted our experiment by dividing them by 4:1 for train and test, respectively. |
| Hardware Specification | Yes | All experiments, including optimization and inference time measurements, are conducted using a single NVIDIA RTX A6000 GPU. |
| Software Dependencies | No | Our framework is basically built with Py Torch [39], but we manually implement the omnidirectional rasterizer using the CUDA kernel. |
| Experiment Setup | Yes | We follow the hyper-parameters of original 3DGS [27] excluding some hyperparameters. Firstly, we set iterations as 200k, densify_until_iter as 100k. For densification we set percent_dense as 1e-3, densify_grad_threshold_min (τmin) as 2e-5, and densify_grad_threshold_max (τmax) as 1e-4. |