GIFT Professor Ruigang Yang's Paper Accepted to ECCV 2026: GeoV2V, a Geometry-Grounded Video Diffusion Model for Driving Scene Generation

The 19th European Conference on Computer Vision (ECCV 2026) was held in Malmö, Sweden, from September 8 to 12, 2026. Held biennially, ECCV is recognized as one of the three top conferences in computer vision, bringing together researchers from universities, research institutions and industry worldwide for academic exchange and the presentation of research outcomes. A paper by Professor Ruigang Yang's research team was accepted to ECCV 2026.

Yuchen Xi, Yuhan Guo (co-first authors), Chenwei Hou, Zixu Liu, Jin Fang, and Jian Liu authored the paper under the supervision of Professor Ruigang Yang at the Global Institute of Future Technology (GIFT), Shanghai Jiao Tong University. Ruigang Yang served as the corresponding author.

GeoV2V: Geometry-Grounded Video Diffusion Model for Driving Scene Generation

Novel view synthesis under trajectory changes is essential for autonomous driving simulation, yet existing methods struggle to generate consistent videos when extrapolating to unseen viewpoints such as lane shifts.

To address the challenge, the researchers developed a geometry-grounded video diffusion framework GeoV2V. The framework synthesizes temporally coherent driving videos along novel trajectories by integrating LiDAR and depth-derived geometric priors with full-video conditioning.

The study also introduces Para4D, a synchronized dataset that captures synchronized multi-lane driving videos within dynamic scenes. It provides explicit supervision for lane-shift viewpoint transformation. Experiments on Waymo, nuScenes, and Para4D show that the method improves visual quality, geometric consistency and dynamic lighting performance under large trajectory shifts, supporting autonomous driving simulation and data generation.

Figure 1: Overview of the GeoV2V Pipeline for lane-shifting novel view synthesis.

Figure 2: Comparison of novel-view synthesis on the Waymo dataset. Results are shown for leftward offsets of 1 m, 2 m, and 4 m; GT denotes the original trajectory viewpoint. Under its official setting, FreeVS generates only the lower half of the image.

Link to the paper:

https://doi.org/10.1007/978-3-032-37044-0_36

Professor Profile

Ruigang Yang is a Chair Professor and doctoral supervisor at the Global Institute of Future Technology, Shanghai Jiao Tong University. He received his Ph.D. degree in Computer Science from North Carolina State University at Chapel Hill in 2003. He previously served as a professor at the University of Kentucky, Chief Scientist of 3D Vision at Baidu, and Chief Technology Officer of the autonomous driving company Inceptio Technology. His research interests include artificial intelligence (AI), computer vision, robotics, autonomous driving and extended reality (XR/VR). He currently serves as an editor of the journal Machine Vision and Applications and as Vice Chair of the CSIG 3D Vision Technical Committee.