Definition
Computer vision is the field of building computational systems that derive measurements, representations, predictions, or actions from images, video, and other visual signals. A vision system turns sensor data into claims about what is present, where it is, how it is changing, or what action should follow.
Common tasks include image classification, object detection, segmentation, tracking, optical character recognition, pose and depth estimation, three-dimensional reconstruction, and visual question answering. A production system may combine several tasks. An autonomous vehicle, for example, needs more than object labels; it must locate objects, track motion, estimate geometry, and connect uncertain observations to control decisions.
The image is part of the system
Performance depends on the camera and environment as well as the model. Resolution, lens distortion, viewpoint, lighting, motion blur, compression, occlusion, and timing can change the input distribution. Ground-truth labels also depend on task rules and annotator visibility. A benchmark on clean images does not establish performance on a different sensor or operating condition.
Modern computer vision uses convolutional networks, transformers, multimodal models, and classical geometric or signal-processing methods. Architecture choice does not define the field.
Distinguish it from nearby terms
- Image processing transforms an image, such as denoising or resizing it. Computer vision usually infers something about the scene or objects represented.
- Computer graphics synthesizes images from models of a scene. Vision often works in the opposite direction, inferring scene properties from images, though the fields share methods.
- Multimodal AI connects visual information with text, audio, actions, or other modalities. Computer vision can operate on visual signals alone.
- Image generation produces visual content. It is not evidence that the system can measure or reason reliably about an observed scene.
Operational significance
Specify the task, sensor, operating conditions, latency, acceptable uncertainty, and error costs. For high-impact use, report performance across the conditions and populations the system will encounter and define what happens when the visual evidence is poor.
Check your understanding
A model identifies pedestrians in daylight photographs. That result does not establish that the deployed system can detect a partially occluded pedestrian at night from a moving camera. Those are different operating conditions and need separate evidence.