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..
Robust Neural Rendering in the Wild with Asymmetric Dual 3D Gaussian Splatting
Authors: CHENGQI LI, Zhihao Shi, Yangdi Lu, Wenbo He, Xiangyu Xu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on challenging real-world datasets demonstrate that our method consistently outperforms existing approaches while achieving high efficiency. See the project website at https://steveli88.github.io/Asym GS. ... We conduct extensive evaluations across a diverse set of in-the-wild 3D scene reconstruction datasets, demonstrating that our method consistently achieves state-of-the-art performance and efficiency, highlighting its robustness and generality. |
| Researcher Affiliation | Academia | 1Mc Master University 2Xi an Jiaotong University |
| Pseudocode | Yes | See Algorithm 1 in the supplementary material for full details of the Multi-Cue Adaptive Mask. ... Algorithm 1 Multi-Cue Adaptive Masking |
| Open Source Code | No | See the project website at https://steveli88.github.io/Asym GS. ... the code will be released upon publication. |
| Open Datasets | Yes | We evaluate our method on three in-the-wild datasets: the Ne RF On-the-go dataset [18], the Robust Ne RF dataset [19], and the Photo Tourism dataset [6]. ... Table 7: The code repo and licenses. Ne RF On-the-go dataset [18] https://github.com/cvg/nerf-on-the-go Apache 2.0 License Robust Ne RF dataset [19] https://robustnerf.github.io/ Custom Photo Tourism dataset [6] https://github.com/ubc-vision/image-matching-benchmark Apache 2.0 License |
| Dataset Splits | Yes | Table 5: In-the-wild 3D reconstruction datasets. Dataset Scene # Train # Test Distractor Appear. change Ne RF On-the-go [18] Patio-high 222 45 ... Robust Ne RF [19] Statue 255 19 ... Photo Tourism [6] Brandenburg Gate 763 10 |
| Hardware Specification | Yes | Experiments were performed on an RTX 4090. This information, along with training times, is provided in the supplementary. |
| Software Dependencies | No | Our base model is built on Mip-Splatting [25]. ... Semantic regions for the multi-cue adaptive mask are generated using Semantic SAM [12] to create instance-level segmentations and apply Algorithm 1 to select distractor regions as masks. |
| Experiment Setup | Yes | We train for 30,000 iterations on Ne RF On-the-go and Robust Ne RF, with densification and pruning every 1,000 steps until iteration 15,000; and for 100,000 iterations on Photo Tourism, with densification and pruning every 1,000 steps until iteration 50,000. We omit the opacity reset and apply a 1,000-step warm-up before the mutual consistency regularization begins. The consistency regularization weight is set to 0.1. The learnable mask is optimized by a loss weighted λmask = 1.0 with a learning rate of 0.1. For EMA, we use a smoothing factor of β = 0.8. ... Table 6: The other 3DGS-related hyperparameters. |