PRISM-1

Overview

PRISM-1 is a scene reconstruction model for 4D scenes (3D in space + time) from video data. It separates static and dynamic elements self-supervised, handles diverse urban elements (vehicles, pedestrians, cyclists, traffic lights, reflections), and enables novel view synthesis and resimulation for autonomous driving simulation within Ghost Gym.

Key features

• ▸4D scene reconstruction (3D space + time) from video data
• ▸Self-supervised separation of static and dynamic elements
• ▸Handles vehicles, pedestrians, cyclists, traffic lights, brake lights, indicator lights, windshield wipers, roadside debris, leaves, reflections
• ▸Novel view synthesis from unseen viewpoints
• ▸Freeze-time and freeze-position camera manipulation
• ▸Implicit scene flow and geometric consistency
• ▸Generalizes across arbitrary camera rigs without additional sensors
• ▸Scalable representation for complex urban scenes
• ▸Resimulation for closed-loop autonomous driving simulation
• ▸Part of Ghost Gym neural simulator