What is Computer Vision?
Computer Vision is the field that develops computational methods for extracting, estimating, or generating task-relevant representations from visual measurements such as images, video, depth, medical scans, or point clouds.
Quick Facts
| Created | 1960s (early research), modern era from 2012 (AlexNet) |
|---|---|
| Specification | Official Specification |
How It Works
The sensor, task, and label space form one contract
Specify the sensor and capture conditions, input resolution and color space, unit of analysis, expected output, label ontology, annotation rules, confidence semantics, and action that consumes the result. Classification labels a whole input, detection localizes instances with boxes, segmentation predicts pixel regions, tracking maintains identities over time, and geometric vision estimates pose, motion, or depth. A result from one task cannot substitute for another.
Model families impose different assumptions
Classical pipelines use calibration, geometry, filters, descriptors, and optimization; CNNs learn spatial features with local operations and weight sharing; Vision Transformers operate over visual Tokens; Vision-Language Models connect visual and language representations. Production systems can combine them. The right choice depends on supervision, operating conditions, latency, memory, interpretability, and failure cost rather than a universal architecture ranking.
Metrics must match the output and threshold
Classification needs per-class precision, recall, calibration, and confusion analysis. Detection requires a matching rule such as Intersection over Union, a confidence threshold, and metrics such as Average Precision and Recall across classes and object sizes. Segmentation uses mask overlap; tracking and pose need their own protocols. Record preprocessing, model revision, evaluation code, dataset version, and all thresholds so a score can be reproduced.
Deployment conditions create silent failures
Lighting, viewpoint, occlusion, motion blur, compression, camera replacement, weather, background, object scale, and population changes can shift performance while confidence remains high. Test deployment-like slices and negative cases, monitor input and outcome drift, preserve difficult examples for review, and define safe fallback behavior. High-impact uses also require privacy, retention, bias, security, human oversight, and incident controls around the visual model.
Key Characteristics
- Maps visual measurements to a declared semantic, geometric, or generative output
- Covers images, video, depth, medical scans, point clouds, and other sensors
- Uses geometry, signal processing, classical features, CNNs, Transformers, or hybrid methods
- Depends on sensor, preprocessing, ontology, annotation, and threshold contracts
- Requires task-specific metrics and deployment-condition slices
- Creates privacy, safety, drift, and human-oversight obligations
Common Use Cases
- Industrial inspection with measured defect recall and a manual-review queue
- Document understanding that links OCR fields to page regions and provenance
- Medical-image decision support within validated modality and population boundaries
- Robotics perception with calibrated sensors and tested fallback behavior
- Accessibility descriptions that retain uncertainty and permit correction
Example
Loading code...Frequently Asked Questions
How is computer vision different from image processing?
Image processing transforms visual data, for example by resizing, denoising, or correcting color. Computer vision estimates information about a scene, object, document, motion, or geometry. Processing can be one stage in a vision system, but a cleaner image does not by itself provide a semantic or geometric result.
What is the difference between classification, detection, and segmentation?
Classification assigns labels to a complete input. Detection predicts object classes and locations, usually as bounding boxes. Semantic segmentation labels pixels by class, while instance segmentation separates individual objects. Because the outputs differ, each task needs its own annotations, matching rules, metrics, and downstream handling.
How do CNNs, Vision Transformers, and VLMs differ?
CNNs use convolutional operations and spatial weight sharing, Vision Transformers represent visual inputs as Tokens processed with attention, and VLMs align or jointly process visual and language representations. They are model families, not complete systems. The right choice depends on the task, data, latency, memory, interpretability, and evaluation results.
How should an object detector be evaluated?
Freeze the dataset, class ontology, box convention, preprocessing, model revision, confidence threshold, and matching rule. Report precision and recall across Intersection over Union thresholds and object sizes, inspect localization and background errors, and test deployment slices such as lighting, occlusion, camera, weather, and rare classes.
Does a high benchmark score prove a vision system is safe in production?
No. A benchmark covers a particular data distribution, label policy, metric, and threshold. Production can differ in sensors, lighting, viewpoints, object prevalence, user population, and downstream consequences. Validate the complete pipeline on deployment-like data, monitor drift, define fallback and review paths, and test failures that matter to the real decision.