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..
IBGS: Image-Based Gaussian Splatting
Authors: Hoang Chuong Nguyen, Wei Mao, Jose M. Alvarez, Miaomiao Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on standard NVS benchmarks show that our method significantly outperforms prior Gaussian Splatting approaches in rendering quality, without increasing the storage footprint. Our project page is available at https://hoangchuongnguyen.github.io/ibgs. |
| Researcher Affiliation | Collaboration | 1Australian National University 2NVIDIA EMAIL EMAIL EMAIL EMAIL |
| Pseudocode | No | The paper describes methods in paragraph form and does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | 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: [No] Justification: Our code will be released upon acceptance. |
| Open Datasets | Yes | Dataset. Following 3DGS [16], we evaluate the NVS performance of our method using 2 scenes in the Tanks and Temples (TNT) [20] dataset, 2 scenes in the Deep Blending [14] dataset, and 9 scenes in the Mip-Ne RF360 [2] dataset. We also show the results on 3 scenes in the Shiny dataset [32] which pose challenging view-dependent effects including specular highlight, reflection and disc diffraction. |
| Dataset Splits | Yes | For all scenes, we use every 8th image for evaluation, and the rest for training. |
| Hardware Specification | Yes | All experiments are conducted using a single RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer [19]' but does not specify a version number for it or any other software libraries or programming languages. |
| Experiment Setup | Yes | Similar to 3DGS [16], we train our method for 30,000 iterations. During the first 7,000 iterations, we set λ1 = λ2 = 0 and only activate the photometric and normal consistency loss thereafter with λ1 = 0.3 and λ2 = 0.03. The weight γ is initially set to 1, and then decreased to 0.5 during the last 20,000 iterations. Regarding hyper-parameters, we set SH degree l = 2, number of median intersection points K = 4, number of candidate source views S = 4, number of visible source views M = 3 and depth error threshold τ = 0.001. We also prune Gaussians with opacity lower than 0.05. Following [6], we apply the exposure compensation from [6] and our proposed exposure correction method only to the TNT dataset. We use Adam optimizer [19] to train the residual prediction network. The initial learning rate is 0.001, which halves at iterations 18,000 and 25,000. |