Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Motion Matters: Compact Gaussian Streaming for Free-Viewpoint Video Reconstruction
Authors: Jiacong Chen, Qingyu Mao, Youneng Bao, Xiandong MENG, Fanyang Meng, Ronggang Wang, Yongsheng Liang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments 4.1 Experimental Setup We evaluate our method on two widely-used public benchmark datasets. (1) Neural 3D Video (N3DV) [1] consists of six indoor video sequences captured by 18 to 21 viewpoints. (2) Meet Room [37] comprises four indoor scenes recorded with a 13 cameras multi-view system... 4.2 Quantitative Comparisons We conduct quantitative comparisons on existing online methods... Tab. 1 shows that our Com GS achieves competitive results among existing online FVV methods... 4.4 Ablation Study To validate the effectiveness of our proposed methods, we ablate three components of Com GS framework in Tab. 3. |
| Researcher Affiliation | Academia | 1College of Applied Technology, Shenzhen University, Shenzhen, China 2College of Big Data and Internet, Shenzhen Technology University, Shenzhen, China 3College of Electronics and Information Engineering, Shenzhen University, Shenzhen, China 4Department of Computer Science, City University of Hong Kong, Hong Kong, China 5Pengcheng Laboratory, Shenzhen, China 6School of Electronic and Computer Engineering, Peking University, Shenzhen, China |
| Pseudocode | No | The paper describes the method using textual explanations and mathematical equations, along with a pipeline diagram (Figure 2), but does not contain any explicit 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | Justification: We would release the code after acceptance. |
| Open Datasets | Yes | We evaluate our method on two widely-used public benchmark datasets. (1) Neural 3D Video (N3DV) [1] consists of six indoor video sequences... (2) Meet Room [37] comprises four indoor scenes... |
| Dataset Splits | Yes | In both of two datasets, we employ the first view for testing... Following the previous works [2, 1, 8], we downsample the videos by a factor of 2 for training and testing and employ the central view for testing view. (2) Meet Room dataset [37]... The center reference camera is also used for testing. |
| Hardware Specification | Yes | Our method is implemented on an NVIDIA A100 GPU. |
| Software Dependencies | No | Our code is based on the open-source code of 3DGStream [2]. On both N3DV and Meet Room dataset, we utilize COLMAP [50] to generate the initial point cloud and vanilla 3DGS [17] to initialize the Gaussians for 3000 epochs at first frame. |
| Experiment Setup | Yes | Our method is implemented on an NVIDIA A100 GPU. We train 150 epochs for non-key frames reconstruction and 1000 epochs for key frames fine-tuning... For the balance of visual quality and storage requirements, we set spherical harmonics (SH) degree to 1. During training, the learning rate for Gaussian attributes is set to 0.002, for the attributes of the adaptive influence region to 0.02, and for the learnable mask mi to 0.01... We set λD SSIM = 0.2 and λerror = 0.001 in this paper. |