This work presents a novel 3D Gaussian Splatting framework that robustly handles moving objects in dynamic street scenes, enabling high-fidelity reconstruction for autonomous driving and urban mapping.

DeepCox Technical Support

• Algorithm Design: Developed a motion-aware regularization loss to separate static and dynamic components.
• GPU Optimization: Implemented custom CUDA kernels to accelerate splatting and differentiable rendering.
• Data Pipeline: Built preprocessing tools for large-scale street-view datasets (e.g., Waymo, nuScenes).
• Evaluation Framework: Created benchmark scripts for PSNR, SSIM, and LPIPS metrics under occlusion.