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..
Unifying Appearance Codes and Bilateral Grids for Driving Scene Gaussian Splatting
Authors: Nan Wang, Lixing Xiao, Yuantao Chen, Weiqing Xiao, Pierre Merriaux, Lei Lei, Ziyang Yan, Saining Zhang, Shaocong Xu, chongjie Ye, Bohan Li, Zhaoxi Chen, Tianfan Xue, Hao Zhao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present a comprehensive quantitative analysis in Tab. 1, comparing our full model against the Omni Re and its variants with appearance codes (w/AC) and a single bilateral grid (w/BG). Our method consistently sets a new state-of-the-art across all four datasets in both appearance and, most critically, geometric metrics. This is evidenced by our evaluation on multiple widely used datasets, including Waymo [30], Nu Scenes [1], Argoverse [40], and Panda Set [42]. |
| Researcher Affiliation | Collaboration | Nan Wang1 , Yuantao Chen1, Lixing Xiao1, Weiqing Xiao1, Bohan Li3,4 Zhaoxi Chen1, Chongjie Ye1, Shaocong Xu1, Saining Zhang1, Ziyang Yan1 Pierre Merriaux6, Lei Lei6, Tianfan Xue5, Hao Zhao1,2 1BAAI; 2AIR, THU; 3SJTU; 4EIT(Ningbo); 5CUHK; 6Leddar Tech |
| Pseudocode | No | The paper describes methods using mathematical formulations and descriptive text, for example in Section 3 Methodology and Appendix A.1 Loss Functions and Optimization, but it does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Our code is open-sourced at https://bigcileng.github.io/bilateral-driving. |
| Open Datasets | Yes | We evaluate on four autonomous driving datasets: Waymo [30], Nu Scenes [1], Argoverse [40], and Panda Set [42] |
| Dataset Splits | Yes | Datasets. We evaluate on four autonomous driving datasets: Waymo [30], Nu Scenes [1], Argoverse [40], and Panda Set [42], selected for their diversity in sensor configurations (Li DAR, camera specifications), environmental conditions (lighting, seasons), and geographic locations. Identical model hyperparameters are used across all datasets to test generalization capability. Technical specifications including camera counts and ego-vehicle view cropping details are provided in Appendix A2. Appendix B provides dataset-specific details regarding our evaluation protocol, including sequence IDs for reproducibility. For example, for Nu Scenes: "We utilize all six available cameras and all Li DAR sensors. We select the following 8 sequences: 152, 164, 171, 200, 209, 359, 529, 916" and "To address ego-vehicle visibility, we crop the bottom 80 pixels from the back camera images." |
| Hardware Specification | Yes | For each dataset, We train all the methods on a single NVIDIA L40 GPU |
| Software Dependencies | No | The paper does not provide specific software dependency versions (e.g., Python, PyTorch, CUDA versions) in the main text or the appendix. It focuses on the methodological details, loss functions, and optimization strategies without detailing the exact software environment. |
| Experiment Setup | Yes | Training. To optimize our multi-scale Gaussian scene representation, we employ a joint training strategy that minimizes a composite reconstruction loss : Lrecon = λr L1 + λs LSSIM + λd Ld + λo Lo , (9) Furthermore, we introduce two regularization terms to enhance image fidelity: Adaptive Total Variation Regularization. This term encourages smoothness and reduces noise while preserving image details, and Circle Regularization Loss. This loss applies inverse transformation to the groundtruth images, preventing discrepancies and image quality degradation. As visually demonstrated in Fig. 5, these terms effectively improve image fidelity. Detailed are provided in Appendix A1. A.1.4 Coarse-to-Fine Optimization Strategy. We employ a coarse-to-fine optimization strategy by utilizing level-dependent learning rates (e.g., 1 10 5, 3 10 5, 1 10 4 from coarse to fine). Coarser grids, assigned higher learning rates, rapidly learn the global scene illumination, while finer grids, with lower learning rates, hierarchically refine high-frequency photometric details. |