ICON Continuous 3D Perception Model with Persistent State

Qianqian Wang1,2*     Yifei Zhang1*     Aleksander Holynski1,2     Alexei A Efros1     Angjoo Kanazawa1
1UC Berkeley      2Google DeepMind




Given a stream of images, our method produces dense scene reconstruction (point clouds and camera parameters) in a feedforward way.


Dynamic Scenes  •  Static Scenes


Following: Viewpoint following the camera
Fixed: Viewpoint fixed in space