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..

VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment

Authors: Qing Li, Huifang Feng, Xun Gong, Yu-Shen Liu

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on standard benchmarks demonstrate that our method achieves stateof-the-art performance in both surface reconstruction and novel view synthesis. The source code is available at https://github.com/Leo QLi/VA-GS.
Researcher Affiliation Academia Qing Li1 Huifang Feng2 Xun Gong1 Yu-Shen Liu3 1 Southwest Jiaotong University, Chengdu, China 2 Xihua University, Chengdu, China 3 Tsinghua University, Beijing, China
Pseudocode No The paper describes the methodology in text and mathematical equations, but does not explicitly contain a pseudocode block or algorithm figure.
Open Source Code Yes The source code is available at https://github.com/Leo QLi/VA-GS.
Open Datasets Yes We evaluate our surface reconstruction performance on the DTU [18] and Tanks and Temples (TNT) [22] datasets. For novel view synthesis, we use the Mip-Ne RF 360 dataset [2]
Dataset Splits Yes Following 3DGS [21], one out of every eight images is used for evaluation, while the remaining seven are used for training.
Hardware Specification Yes All experiments are conducted on a single NVIDIA RTX 4090 GPU.
Software Dependencies No The paper mentions COLMAP for initializing 3D Gaussians but does not specify software names with version numbers for key libraries or environments used in the experiments.
Experiment Setup Yes We set the number of source views to N = 3, the threshold in Lns to τ = 0.01, and the patch size in Lp to 7 7. The loss weight factors are set as follows: β1 = 0.2, β2 =0.03, λ1 =0.015, λ2 =0.3, λ3 =0.15, and λ4 =1.0. The model is trained for 20,000 iterations for surface reconstruction and 30,000 iterations for novel view synthesis. We first pretrain the model using only the color loss for 7,000 steps to obtain a coarse geometric initialization, which provides a stable foundation for subsequent geometry refinement. Then, we incorporate our image edge item and normal-based geometry alignment into the training. To further refine geometry, we sequentially apply our multi-view photometric alignment for 8,000 iterations, followed by 5,000 iterations of multi-view feature alignment. For novel view synthesis, we continue training for an additional 10,000 steps to optimize rendering quality.