What is Computer Vision?
Computer vision is the field concerned with extracting information from images and video so that software can describe or act on visual content. Typical tasks include classifying an image, detecting and locating objects, segmenting regions at the pixel level, tracking motion across frames, and reconstructing three dimensional structure. Modern systems are dominated by learned models rather than hand written rules.
Before deep learning, pipelines relied on hand designed feature detectors that responded to corners, edges, and gradients, followed by a classifier. Convolutional neural networks replaced most of that design work by learning filters directly from labeled images, and transformer based architectures later showed that image patches can be modeled with the same attention mechanisms used for text.
Data and evaluation shape the field heavily. Large annotated image collections made supervised training possible, and standardized benchmarks made progress comparable. Reported accuracy is tied to the conditions represented in those datasets, so lighting, camera, viewpoint, resolution, and demographic coverage all affect whether laboratory performance transfers to a specific deployment. Auditing accuracy separately across relevant subgroups is standard practice.
Known weaknesses include sensitivity to small input perturbations that do not affect human perception, degradation under distribution shift such as weather or new equipment, and uneven error rates across groups when training data is unbalanced. These properties are especially consequential in identification, medical, and safety applications, where documentation and human review are commonly required.
Vision is increasingly combined with language. Multimodal models accept images alongside text and answer questions about them, caption them, or reason over diagrams, and the same embedding techniques used for text retrieval are applied to images for visual search. Video adds a temporal dimension, and with it action recognition and considerably higher computational cost.
Key points
- Extracts structured information from images and video.
- Tasks include classification, detection, segmentation, and tracking.
- Convolutional and transformer models replaced hand designed features.
- Accuracy depends heavily on dataset coverage and conditions.
- Increasingly merged with language in multimodal models.
In practice
A quality inspection camera photographs each part on a line. A segmentation model marks the pixels belonging to the component, and a detector flags scratches longer than a threshold. Flagged parts are diverted for human inspection. When the plant changes its lighting, false positives rise, and the team retrains on freshly labeled images captured under the new conditions.