Computer vision
Computer vision is the field concerned with extracting meaning from images and video: where objects are, what they are, and how they move. Face recognition, video analytics and optical character recognition are all applications of it, built on the same underlying step of turning pixels into structured measurements.
What are the core computer vision tasks?
- Classification: what is in this image, as a single label.
- Detection: where each object is, as a box with a class.
- Segmentation: which pixels belong to which object, rather than a rectangle around it.
- Tracking: keeping the same identity on an object across frames of a video.
- Recognition: matching an instance to a known one, which is what face recognition performs.
- Pose and keypoint estimation: where the parts of an object or body are relative to each other.
Applied systems chain these rather than choosing one. A face recognition pipeline detects, estimates keypoints to align, then recognises; a video analytics pipeline detects, tracks, then applies a rule. The chain is why a weakness early on cannot be recovered later: a face never detected is never recognised.
How do face recognition and video analytics relate to computer vision?
Both are applications of computer vision distinguished by what they output. Video analytics stops at classes and movement: a person, a vehicle, a zone entry. Face recognition continues to identity by comparing a face against enrolled templates. The shared foundation is detection, which is why one platform commonly offers both and why buyers meet the terms together.
The legal treatment does not follow the shared foundation. Identifying a person is biometric processing in most jurisdictions; counting people is not. Two features from one platform can therefore sit in entirely different regulatory categories.
What limits computer vision accuracy in the field?
| Factor | Effect | What helps |
|---|---|---|
| Pixels on target | Detection fails before recognition does | Camera placement and focal length for the actual distance |
| Lighting and backlight | Loss of the texture the model relies on | Positioning away from direct sun, and WDR-capable cameras |
| Pose and occlusion | Partial views score lower against enrolled templates | A view angle close to how people naturally pass |
| Motion blur | Frames become unusable at a walking pace | Shutter speed appropriate to the traffic, not to the scene brightness |
| Domain shift | A model trained elsewhere behaves differently here | Evaluation on footage from this site before procurement |
Every one of these is a capture decision rather than a software one, which is the practical reason a pilot on the site's own cameras predicts a deployment better than any benchmark. The benchmark measures the model; the pilot measures the installation.
Frequently asked questions
- What is computer vision?
- Computer vision is the field concerned with extracting meaning from images and video: where objects are, what they are and how they move. Face recognition, video analytics and optical character recognition are applications of it, all built on turning pixels into structured measurements.
- Is face recognition part of computer vision?
- Yes. Face recognition is a computer vision application that continues past detection to identity, matching a detected face against enrolled templates. It shares its detection and alignment stages with other vision tasks and differs in what it outputs and in how it is regulated.
- What is the difference between computer vision and machine learning?
- Machine learning is the method; computer vision is the problem domain. Most modern vision systems are built with machine learning, but the field also includes geometric and optical techniques that involve no learning at all, such as camera calibration and classical stereo reconstruction.
- What limits computer vision accuracy in real deployments?
- Capture conditions, more than model choice: pixels on the target, lighting and backlight, pose and occlusion, motion blur, and the difference between the footage a model was trained on and the site's own. These are placement and optics decisions, which is why a pilot predicts results better than a benchmark.
- Does computer vision require the cloud?
- No. Vision models run on local hardware routinely, including on cameras and small appliances, and for multi-camera video the local option is usually the practical one because sending every stream to a remote service is a continuous bandwidth commitment.
- What are the main computer vision tasks?
- Classification labels an image, detection locates objects, segmentation assigns pixels to objects, tracking maintains identity across frames, recognition matches an instance to a known one, and pose estimation locates parts relative to each other. Applied systems chain several of these rather than using one.