Computer Vision
Seeing is not the same as labeling. Our computer vision research studies how visual systems recover structure — objects, geometry, motion, and relationships — and how those representations support reasoning about the physical world.
- Visual representation learning
- Image understanding
- Object recognition
- Video understanding
- Spatial reasoning
- 3D perception
- Vision–language learning
Why it matters
Recognition benchmarks have largely saturated, yet visual systems still struggle with tasks that people find trivial: counting, judging relative position, tracking an object behind an occluder, or understanding that a scene seen from two angles is the same scene. These gaps matter for any application where a system must act on what it sees.
Vision is also an unusually rich source of supervision. Images and video contain structure — geometry, physics, causality — that can be learned without labels. Understanding how to extract that structure is central to building systems that learn about the world rather than about datasets.
Key research questions
What representations allow a model to reason about spatial relationships rather than recognize them by appearance?
How much of three-dimensional structure can be recovered from unposed images and video alone?
Why do vision–language models fail on compositional queries, and what training signals correct this?
How can temporal models capture cause and effect in video rather than correlation between frames?
What makes a visual representation robust to changes in viewpoint, lighting, and context?
Primary subfields
- Visual representation learning
- Self-supervised and supervised features for images and video.
- 3D perception
- Depth, geometry, and scene reconstruction from visual input.
- Video understanding
- Motion, events, and temporal structure.
- Spatial reasoning
- Relations, counting, and geometric inference.
- Vision–language learning
- Aligning visual and linguistic representations.
Representative methodologies
Synthetic scene generation
Rendered environments with exact ground truth for geometry and relations.
Feature visualization
Inspecting what intermediate layers respond to.
Diagnostic benchmarks
Small, controlled test sets that isolate specific perceptual skills.
Geometric priors
Incorporating projective and physical constraints into learned models.
Relationships to other domains
- Natural Language Processing
Multimodal models connect perception to language.
- Generative AI
Generative image models share — and stress-test — visual representations.
- Scientific Discovery
Microscopy, astronomy, and remote sensing are visual sciences.
Working on questions in Computer Vision?
We welcome conversations with researchers and organizations working on related questions — whether to compare notes, discuss a collaboration, or share work in progress.