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..
4DGCPro: Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming
Authors: Zihan Zheng, Zhenlong Wu, Houqiang Zhong, Yuan Tian, Ning Cao, Lan Xu, Jiangchao Yao, Xiaoyun Zhang, Qiang Hu, Wenjun Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that 4DGCPro enables flexible quality and multiple bitrate within a single model, achieving real-time decoding and rendering on mobile devices while outperforming existing methods in RD performance across multiple datasets. 4 Experiments To comprehensively evaluate our method, we conducted experiments on two kind of distinct datasets: (1) the N3DV dataset [31] featuring subtle motions with background, and (2) the Hi Fi4G dataset [26] containing complex motions without background. We additionally captured a new dataset using 81 synchronized Z-CAM cinema cameras 3840 2160, recording diverse performances including dance, sports, and instrument playing. |
| Researcher Affiliation | Collaboration | Zihan Zheng1 , Zhenlong Wu1 , Houqiang Zhong2 , Yuan Tian2,3 , Ning Cao4 , Lan Xu5 , Jiangchao Yao1 , Xiaoyun Zhang1 , Qiang Hu1 , Wenjun Zhang1,2 Cooperative Medianet Innovation Center, Shanghai Jiaotong University1 Department of Electronics, Shanghai Jiatong University2 Shanghai AI Lab3 Cloud platform department, E-surfing Vision Technology Co., Ltd.4 School of Information Science and Technology, Shanghai Tech University5 |
| Pseudocode | No | No explicit pseudocode or algorithm blocks are present in the paper. The methodology is described in prose with mathematical formulations. |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The paper provides open access to both the data and the code. |
| Open Datasets | Yes | To comprehensively evaluate our method, we conducted experiments on two kind of distinct datasets: (1) the N3DV dataset [31] featuring subtle motions with background, and (2) the Hi Fi4G dataset [26] containing complex motions without background. We additionally captured a new dataset using 81 synchronized Z-CAM cinema cameras 3840 2160, recording diverse performances including dance, sports, and instrument playing. The datasets used in the research are all publicly available standard datasets and self-made datasets that will be made public, which do not contain any data involving personal privacy or sensitive information. |
| Dataset Splits | Yes | We selected the 48th view as the test view in the Hi Fi4G and 4DGCPro datasets, while the 0th view was chosen in the N3DV dataset. |
| Hardware Specification | Yes | Our experimental setup includes an Intel(R) Xeon(R) W-2245 CPU running at 3.90 GHz and an RTX 3090 graphics card. |
| Software Dependencies | No | The H.264 encoder was configured using the x264 library with the following settings: I/P-frames only (no B-frames), 3 reference frames, color space in YUV4:4:4, and preset set to "medium." (No version numbers provided for H.264 encoder or x264 library). |
| Experiment Setup | Yes | During the pre-training phase of key frames, we first train for 12,000 steps under the setting of λssim = 0.2. In the Gaussian pruning phase, we remove 40% of the Gaussians with lower opacity on Hi Fi4G dataset [26] and 4DGCPro dataset but not remove any Gaussians on N3DV dataset [31]. During the hierarchical process, we set λΨ = 1 105 to ensure a balance between volume and opacity, and divide the Gaussians into L = 6 layers. Regarding the motion-aware adaptive Gaussian grouping, we have selected different τµ values for different datasets: τµ = 0.0025 for 4DGCpro, τµ = 0.001 for Hi Fi4G, and τµ = 0.01 for N3DV. Then, we conduct end-to-end entropy-optimized training on keyframes for 1,500 steps with the supervision of λl = 0.5/l, l < L 1, l = L, and λrate_key = 1 10 7. In the subsequent inter-frame Optimization., we set λrate_inter = 1 10 4, λreg = 1 10 3, and train for 800 and 2,000 steps in the rigid transformation and residual deformation phases respectively. |