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..
Metropolis-Hastings Sampling for 3D Gaussian Reconstruction
Authors: Hyunjin Kim, Haebeom Jung, Jaesik Park
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on benchmark datasets, including Mip-Ne RF360, Tanks and Temples and Deep Blending, show that our approach reduces the number of Gaussians needed, achieving faster convergence while matching or modestly surpassing the view-synthesis quality of state-of-the-art models. Our project page is available at https://hjhyunjinkim.github.io/MH-3DGS. |
| Researcher Affiliation | Academia | 1UC San Diego 2Seoul National University EMAIL EMAIL |
| Pseudocode | Yes | The complete algorithmic flow is in Algorithm 1 and 2 in Appendix D. |
| Open Source Code | Yes | Our project page is available at https://hjhyunjinkim.github.io/MH-3DGS. |
| Open Datasets | Yes | We demonstrate the effectiveness of our approach on real-world scenes from the Mip-Ne RF360 [1], Tanks and Temples [23], and Deep Blending [17] datasets. |
| Dataset Splits | Yes | We perform comprehensive experiments using the same real-world datasets employed in 3DGS. Specifically, we utilize the scene-scale view synthesis dataset from Mip Ne RF360 [1]... Additionally, we selected two scenes each from the Tanks and Temples [23] and the Deep Blending [17] datasets, using the same scenes as in the original 3DGS. |
| Hardware Specification | Yes | We conducted experiments on an NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | Our framework is based on 3DGS [20] implementation. We also use the differentiable tile rasterizer implementation provided by 3DGS-MCMC [22]. |
| Experiment Setup | Yes | We use λ = 0.2, λopacity = 0.01 and λscale = 0.01 in our framework. ... We use α = 0.8, β = 0.5, γ = 0.5 in our tests. ... During densification, a normalized progress value ( [0, 1]) linearly transitions parameters from coarse (larger offsets, voxel sizes, and batch sizes) to fine (smaller offsets and voxel sizes) to shift focus from broad to precise refinements. Coarse proposal scales range from 10.0 to 5.0, fine proposal scales from 2.0 to 1.0, and voxel sizes from 0.02 to 0.005. Batch sizes for coarse and fine proposals are set to 4,500 and 16,000, respectively. |