A team of Oxford Engineering and Google DeepMind researchers has received the prestigious Best Paper Award at CVPR 2026 for their work Efficiently Reconstructing Dynamic Scenes One D4RT at a Time. This award recognises the top publication out of more than 16,000 submissions to the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) - the principal international conference on computer vision.
The paper was authored by researchers who are current or former members of the Oxford Visual Geometry Group (VGG), together with researchers at Google DeepMind. The VGG researchers are Chuhan Zhang, Liliane Momeni, and Junyu Xie (a current DPhil student), together with Andrew Zisserman (a Royal Society Professor in Engineering Science).
The winning paper introduces D4RT, a novel approach for 3D reconstruction from a video of a dynamic scene using deep neural networks. It can predict a scene’s full 3D geometry, including the cameras, scene depth, and correspondences across the video frames. Unlike previous approaches, D4RT can make predictions for dynamic scenes with moving objects, as well as static scenes. It can carry out these multiple tasks from a video in just a few seconds.
