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..
DC4GS: Directional Consistency-Driven Adaptive Density Control for 3D Gaussian Splatting
Authors: Moonsoo Jeong, Dongbeen Kim, Minseong Kim, Sungkil Lee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our DC4GS on the three standard datasets: Mip-Ne RF 360 [2] (five outdoor and four indoor scenes), and two scenes each from Tanks and Temples [16] and Deep Blending [12], following the 3DGS [14]. Every 8th frame is used for test. Rendering quality is evaluated by PSNR, SSIM [35], and LPIPS [43]. We also report memory requirements for 3DGS parameters and the number of primitives to assess storage efficiency. |
| Researcher Affiliation | Academia | Moonsoo Jeong1 Dongbeen Kim2 Minseong Kim3 Sungkil Lee1,2,3, 1Department of Electrical and Computer Engineering, Sungkyunkwan University, South Korea 2Department of Computer Science and Engineering, Sungkyunkwan University, South Korea 3Department of Immersive Media Engineering, Sungkyunkwan University, South Korea EMAIL |
| Pseudocode | Yes | Algorithm 1 Optimization and Densification Algorithm 2 Directional consistency (DC) from positional gradients Algorithm 3 DC-based split costs for N (odd) candidates along the principal axis. Algorithm 4 DC-based cost for a candidate split xk Algorithm 5 Splits a Gaussian into sub-primitives at xopt |
| Open Source Code | Yes | Corresponding author. Code is available at https://github.com/cgskku/dc4gs. |
| Open Datasets | Yes | We evaluate our DC4GS on the three standard datasets: Mip-Ne RF 360 [2] (five outdoor and four indoor scenes), and two scenes each from Tanks and Temples [16] and Deep Blending [12], following the 3DGS [14]. And from Appendix A.6: Mip-Ne RF360 [2]: no explicit license terms provided. Available at https://jonbarron.info/mipnerf360/. Tanks and Temples [16]: released under the Creative Commons Attribution 4.0 International (CC BY 4.0). Available at https://www.tanksandtemples.org/license/. Deep Blending [12]: no explicit license terms provided. Available at http://visual.cs.ucl.ac.uk/pubs/deepblending/. |
| Dataset Splits | Yes | Every 8th frame is used for test. |
| Hardware Specification | Yes | All the experiments are conducted on a single NVIDIA A6000 GPU with 48GB of memory. |
| Software Dependencies | No | Our DC4GS can be integrated seamlessly into the existing 3DGS pipelines. We implemented ours on top of the 3DGS [14], as well as its recent extensions: Pixel-GS [44] and Abs GS [38]. |
| Experiment Setup | Yes | To ensure a fair comparison with the baselines [14, 38, 44], we adopt their training schedules, loss functions, and all hyperparameters, including the gradient threshold τp and the scale threshold τS. As with the baselines, we halt the densification after 15K iterations. In our implementation, we empirically set the samples of N=5. |