D-MiSo: Editing Dynamic 3D Scenes using Multi-Gaussians Soup
Authors: Joanna Waczynska, Piotr Borycki, Joanna Kaleta, Slawomir Tadeja, Przemysław Spurek
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental section is divided into two parts. First, we show that D-Mi So can model dynamic scenes with high quality; in the second part, we demonstrate that our model allows an easy editing procedure, which is our main contribution. Table 1: Quantitative comparisons (PSNR) on a D-Ne RF dataset showing that D-Mi So gives comparable results with other models. |
| Researcher Affiliation | Academia | Joanna Waczy nska Doctoral School of Exact and Natural Sciences Jagiellonian University, Piotr Borycki* Faculty of Mathematics and Computer Science Jagiellonian University, Joanna Kaleta Warsaw University of Technology Sano Centre for Computational Medicine, Sławomir Tadeja Department of Engineering University of Cambridge, Przemysław Spurek Jagiellonian University IDEAS NCBR |
| Pseudocode | No | The paper describes the method conceptually but does not provide pseudocode or an algorithm block. |
| Open Source Code | Yes | The source code is available on 3. Our code is developed on top of the GS vanilla code, according to their license. (Footnote 3 points to https://github.com/waczjoan/D-Mi So) |
| Open Datasets | Yes | D-Ne RF Datasets: Contains seven dynamic objects with realistic materials described with a single camera [11]. Ne RF-DS[44]: This dataset contains again seven real-world scenarios containing a moving or deforming specular object. Panoptic Sports Datasets: The dataset comprises six dynamic scenes featuring significant object and actor movements [26, 45]. |
| Dataset Splits | No | The paper mentions "training views" and defines "training" and "testing" splits for datasets (e.g., "27 cameras for training and the remaining 4 cameras for testing" for Panoptic Sports), but does not explicitly describe a separate "validation" dataset split or methodology. |
| Hardware Specification | Yes | We used NVIDIA Ge Force RTX 4090 and A100 GPUs. |
| Software Dependencies | No | The paper states "Our code is developed on top of the GS vanilla code" but does not provide specific version numbers for this or any other software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | In our method, a batch of images is taken as input to the model. Tab. 6 shows a numerical comparison using batch: 4, 8 on D-Ne RF. Training takes 80 thousand iterations, and the second stage starts at the 5 thousandth iteration. Each Core-Gaussian has attached 25 Sub-Gaussians. |