Fully Explicit Dynamic Gaussian Splatting
Authors: Junoh Lee, Changyeon Won, Hyunjun Jung, Inhwan Bae, Hae-Gon Jeon
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on various scenes demonstrate the state-of-the-art rendering quality from our method, achieving fast rendering of 62 fps on a single 2080Ti GPU. |
| Researcher Affiliation | Academia | Junoh Lee1, Changyeon Won2, Hyunjun Jung2, Inhwan Bae2, Hae-Gon Jeon1,2 1School of Electrical Engineering and Computer Science 2AI Graduate School Gwangju Institute of Science and Technology {juno,cywon1997,hyunjun.jung,inhwanbae}@gm.gist.ac.kr, haegonj@gist.ac.kr |
| Pseudocode | No | The paper describes the proposed methods in detail but does not include formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide the source code in the supplemental materials along with its instructions. |
| Open Datasets | Yes | The effectiveness of our method is validated through experiments on two major real-world video datasets: Neural 3D Video dataset [7] and Technicolor dataset [20]. |
| Dataset Splits | Yes | We follow a conventional evaluation protocol in [86, 15], which uses subsequences divided from whole videos. We report PSNR, SSIM and LPIPS values. For a fair comparison, we train all models for all 300 frames including concurrent works, except for STG [15], Ne RFPlayer [72] and Hyper Reel [86]. For Ne RFPlayer and Hyper Reel, we directly borrow the results from [72, 86]. For STG, it is not possible to train for all 300 frames due to a GPU memory issue, so we report the results for only 150 frames, which is the maximum duration running on a single NVIDIA H100 80GB GPU. |
| Hardware Specification | Yes | Comprehensive experiments on various scenes demonstrate the state-of-the-art rendering quality from our method, achieving fast rendering of 62 fps on a single 2080Ti GPU. |
| Software Dependencies | No | The paper mentions software like 3DGS [13], Mip-Splatting [83], COLMAP [84], RAdam optimizer [85], and scikit_image library, but it does not specify version numbers for these dependencies. |
| Experiment Setup | Yes | Our codebase is built upon 3DGS [13] and Mip-Splatting [83] and uses almost its hyperparameters. For initialization, our experiments use only COLMAP [84] point clouds from the first frame. The time interval and initial duration are both set to 10. We increment the duration by its interval every 400 iterations. Both static and dynamic regularization parameters are set to 0.0001. We employ the RAdam optimizer [85] for training. |