D4RT: Teaching AI to see the world in four dimensions

Google DeepMind · 6 months ago

Researchers introduced D4RT, an AI model that reconstructs dynamic 3D scenes from 2D video by tracking objects through space and time as a unified system rather than using separate specialized models. The model processes one-minute videos in roughly five seconds on a single TPU chip, performing 18x to 300x faster than previous methods that required up to ten minutes for the same task. This efficiency enables real-time applications in robotics and augmented reality that previously required too much computational power for practical use.