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..
Holistic Large-Scale Scene Reconstruction via Mixed Gaussian Splatting
Authors: Chuandong Liu, Huijiao Wang, Lei YU, Gui-Song Xia
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To comprehensively evaluate our approach, we conduct extensive experiments on two prominent benchmarks [31, 53] for large-scale 3D scene reconstruction, encompassing four challenging scenes. Our results demonstrate that Mix GS significantly improves reconstruction performance in these scenarios. To sum up, our key contributions are as follow: |
| Researcher Affiliation | Academia | 1School of Computer Science, Wuhan University 2School of Artificial Intelligence, Wuhan University 3School of Electronic Information, Wuhan University 4State Key Lab. of LIESMARS, Wuhan University 5Institute for Math & AI, Wuhan University |
| Pseudocode | No | The paper describes the method's steps within the main text and figures (e.g., Figure 2 for pipeline overview), but it does not include a clearly labeled pseudocode block or algorithm. |
| Open Source Code | Yes | Project Page: mixgs.github.io. In the NeurIPS Paper Checklist, Question 5 states: "The code and data will be included in the project page." |
| Open Datasets | Yes | Following state-of-the-art methods [10, 34], we conduct experiments on large-scale scenes across two real-world urban scene datasets: Urban Scene3D [31] dataset with Residence and Sci-Art, and Mill19 [53] dataset with Building and Rubble. |
| Dataset Splits | Yes | For fair comparison [10, 53, 30], we adopt the same data splits as previous works to construct the training and testing sets. Specifically, the Building, Rubble, Residence, and Sci-Art scenes contain 1920, 1657, 2582, and 3620 training images, respectively, and 20, 21, 21, and 21 testing images. |
| Hardware Specification | Yes | All experiments are conducted on an NVIDIA RTX 3090 GPU with 24GB VRAM using PyTorch. For 3DGS and our Mix GS, we conduct experiments on a single RTX 3090 GPU paired with an Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz, and report the corresponding metrics. |
| Software Dependencies | Yes | The software environment includes PyTorch version 2.0.1 and CUDA version 11.7. |
| Experiment Setup | Yes | The three training stages are run for 30,000, 40,000, and 260,000 iterations, respectively. To fit the original 3DGS [22] for large-scale scene reconstruction, we train for 60,000 iterations, applying densification every 200 steps until 30,000 iterations. Following previous methods [34, 10], we use the official camera poses provided by Mega-Ne RF [53] and initialize 3D Gaussians with COLMAP [48]. To ensure a fair comparison, we downsample all images by 4 times. |