Stanford Vision and Learning Lab

Stanford University

How can computational models enable machines to form human-like visual interpretations of objects, scenes, and actions? SVL constructs large-scale datasets and benchmarks such as BEHAVIOR to evaluate embodied AI agents on household tasks in realistic simulation environments. The lab develops algorithms for object and activity recognition including the Multi-Object Multi-Actor (MOMA) framework and activity graphs for compositional activity parsing in videos. SVL integrates multisensory datasets like ObjectFolder with benchmarks to train models that reason about 3D object properties from sight, sound, and touch. Research combines computational geometry, automated image and video analysis, and visual reasoning methods to study foundational principles and practical implementations of high-level visual perception.