DrivingScene: A Multi-Task Online Feed-Forward 3D Gaussian Splatting Method for Dynamic Driving Scenes

Modern autonomous vehicles are typically equipped with multiple cameras for 360-degree surround-view perception. Compared to fusion-based approaches that rely on multi-modal sensors like LiDAR or RaDAR[1, 2, 3], vision-only methods[4, 5] offer a more cost-effective and computationally efficient pathway for complex online perception tasks. However, reconstructing a large-scale, geometrically accurate, and photorealistic dynamic scene in real-time, solely from sparse and dynamic surround-view images, remains a significant and unresolved challenge. The pursuit of higher reconstruction fidelity has seen tremendous success with neural rendering techniques like NeRF [6] and 3DGS [7]. However, the majority of these methods, whether for static scenes like StreetGaussian [8], DrivingGaussian [9] or dynamic scenes like EmerNeRF [10], are bound by a per-scene optimization paradigm. This reliance on time-consuming offline training is incompatible with the real-time requirements of autonomous driving downstream tasks, necessitating a paradigm shift towards "feed-forward" reconstruction[11, 12, 13, 14]. This online approach has matured for static scenes, with methods like pixelSplat [15] and MVSplat [16] demonstrating its viability, and culminating in works like DrivingForward [17] which successfully handle sparse driving contexts. Y et, their foundational static world assumption inevitably leads to severe artifacts when confronted with moving vehicles.