SCOPE: New Deep-Learning Method Achieves Temporally Consistent 3D Geometry from Monocular Video

Researchers from the University of Hong Kong, Alibaba Group, Horizon Robotics, and Ant Group have introduced SCOPE (Scale-Consistent One-Pass Estimation of 3D Geometry), a deep-learning approach for deriving accurate, temporally stable 3D geometry from extended monocular video sequences. Accepted to SIGGRAPH 2026, SCOPE addresses a persistent weakness in existing methods: maintaining both geometric accuracy and temporal consistency across hundreds of frames.

The system introduces three core innovations — viewpoint-invariant geometry, appearance-invariant learning, and frequency-modulated positioning — to produce scale-invariant 3D point maps in a single pass over an entire sequence. In benchmark tests on ScanNet, SCOPE reduced relative point map error by 24.2% and temporal alignment error by 34.9% compared to state-of-the-art methods, with demonstrated capability on complex camera trajectories, variable lighting, portrait video, and 4D scene reconstruction.