Paper accepted to CVPR 2026
Our paper has been accepted for publication at the Conference on Computer Vision and Pattern Recognition (CVPR 2026) Findings Track.
MADrive: Memory-Augmented Driving Scene Modeling by Polina Karpikova, Daniil Selikhanovych, Kirill Struminsky, Ruslan Musaev, Maria Golitsyna, Dmitry Baranchuk
Recent advances in scene reconstruction enable highly realistic modeling of autonomous driving environments using techniques such as 3D Gaussian splatting. However, these reconstructions remain closely tied to the original observations and often struggle to support photorealistic synthesis of significantly modified or entirely new driving scenarios.
In this work, we introduce MADrive, a memory-augmented driving simulation framework that extends existing scene modeling approaches by replacing observed vehicles with visually similar 3D assets retrieved from a large external memory bank. We also introduce MAD-CARS, a curated dataset of about 70K 360° car videos collected in real-world conditions.
The proposed retrieval module identifies the most similar vehicles in the memory bank, reconstructs their 3D representations from video, and integrates them into the target scene using orientation alignment and relighting. As a result, the system provides complete multi-view vehicle representations and enables photorealistic synthesis of substantially modified driving scenes.