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..
FHGS: Feature-Homogenized Gaussian Splatting
Authors: qigeng duan, Benyun ZHAO, Mingqiao Han, Yijun Huang, Ben Chen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive comparison experiments with other state-of-the-art methods on benchmark datasets demonstrate that our FHGS exhibits superior reconstruction performance in feature fusion, noise suppression, and geometric precision, while maintaining a significantly lower training time. |
| Researcher Affiliation | Academia | Q. G. Duan Benyun Zhao Mingqiao Han Yijun Huang Ben M. Chen Department of Mechanical and Automation Engineering The Chinese University of Hong Kong |
| Pseudocode | Yes | Algorithm 1 General Feature Fusion and Densification Framework Algorithm 2 Non-Differentiable Feature Driving Mechanism |
| Open Source Code | Yes | Our code and additional results are available on our project page: https://fhgs.cuastro.org/. |
| Open Datasets | Yes | We conduct systematic experiments on a range of public datasets covering both indoor and outdoor environments. For indoor evaluations, we evaluate our method on DTU (scans 24, 37) [29] and Mip-Ne RF 360 (Kitchen) [15], while outdoor evaluations are performed on Mip-Ne RF 360 (Garden, Stump) and Tanks and Temples (Tn T Caterpillar) [30]. |
| Dataset Splits | No | All input images are uniformly downsampled to a maximum side length of 1,000 pixels to balance computational efficiency and reconstruction accuracy. Sparse point clouds are initialized with COLMAP [31] are used for Sf M, with a fixed iteration count of 10,000 to ensure optimization consistency. During testing, Feature3DGS [10] failed in Tn T [30] due to its huge utilization of GPU memory. The ablation study is conducted on the scan24 of DTU dataset [29] with 10,000 training iterations to investigate the effects of the loss functions Lgt and Lcf on feature fusion, geometric reconstruction, and optimization efficiency (illustrated in Table 4). |
| Hardware Specification | Yes | All experiments are executed on a workstation equipped with a single NVIDIA Ge Force RTX 4090 (24 GB) and an AMD Ryzen 9 9950X (16 cores). |
| Software Dependencies | No | We implement FHGS within a 2DGS-based framework, deploying tailor-made CUDA kernels to accelerate the proposed feature-fusion operations. The original 2DGS renderer is retained to export depth-distortion maps, depth maps, normal maps, and mesh reconstructions, which serve as the inputs to our quantitative and qualitative evaluations. |
| Experiment Setup | Yes | In the sigmoid activation function, the similarity threshold and slope are empirically fixed to λ = 0.5 and k = 20, respectively, ensuring stable binarization of the feature-matching score σ that governs the polarity of gaussian primitive. For benchmarking, we adopt Feature3DGS [10] as the baseline. Following its protocol, we report the L1 feature loss FL1 (lower values indicating higher feature similarity) under the same rendering pipeline, where smaller FL1 values signify better feature fusion. Cross-view consistency is further assessed with the ground-truth entropy metric Lgt (Eq. 5); lower Lgt scores indicate tighter multi-view alignment. ... In addition, we use identical feature-extraction pipelines together with the default 2DGS optimizer settings (learning rate, iteration count, batch size) for both the baseline and our method, thereby eliminating performance biases due to hyperparameter tuning or feature-generation differences. The ablation study is conducted on the scan24 of DTU dataset [29] with 10,000 training iterations to investigate the effects of the loss functions Lgt and Lcf on feature fusion, geometric reconstruction, and optimization efficiency (illustrated in Table 4). |