P2GS：基于物理先验的高斯溅射用于光谱一致的城市重建

3D 高斯溅射 (3DGS) 最近涌现为一种 powerful explicit representation， enabling fast, high-fidelity rendering，使其成为 closed-loop simulator 和 perception models 在自动驾驶中的有前景的基础。然而，conventional 3DGS 隐式 假设 views 间 exposure 和 tone mapping consistent。Real driving data 违反了这一假设 due to heterogeneous camera pipelines 和 dynamic outdoor illumination，将 exposure differences 和 sensor noise baking 到 radiance field 中，producing artifacts and inconsistent illumination，尤其在 static backgrounds 上至关重要用于 realistic simulation。这些问题在自动驾驶中被放大，sparse viewpoints、varying exposures 和 outdoor lighting 相互作用，而 prior work 主要针对 dynamic-object reconstruction，overlooks cross-view photometric consistency。为此，我们引入 P2GS，一个 physically consistent 的高斯溅射框架，jointly 分解 view-invariant linear HDR radiance field、per-view exposure scales 和 tone-mapping functions 从 only LDR images without HDR supervision。P2GS employs a unified optimization strategy grounded in the physical image-formation process，enforcing relative-exposure consistency and HDR-domain radiance regularization。这 yield 一个 robust 于 inter-camera illumination differences 的 radiance field，同时 preserving standard 3DGS 的 real-time efficiency。Experiments across real and simulated driving environments show that P2GS matches or surpasses prior methods in LDR reconstruction while providing substantially improved photometric consistency, reliable exposure normalization, and physically coherent illumination across diverse scenes。