CV
How it works
A typical computer vision pipeline covers image acquisition, preprocessing (normalization, denoising, geometric correction), feature extraction and a decision model. In the classical approach features were hand-engineered (edges, SIFT/HOG descriptors) and classification was performed by machine-learning algorithms. In the modern deep-learning approach convolutional neural networks (CNNs) and vision transformers (ViTs) learn hierarchical representations directly from data: early layers detect simple patterns (edges, textures) and deeper layers detect complex objects. At the end of the pipeline a task head turns the representation into an output: a class label, detection boxes, a segmentation mask or a depth map. Models are trained on large labeled datasets (e.g. ImageNet), optimizing a loss function via gradient descent.
Problem solved
Machines operate on raw pixel arrays that carry no semantic meaning on their own. Computer vision solves the problem of turning this unstructured visual data into useful descriptions of the world — identifying objects, their locations and relations — which is a prerequisite for automating perception in robotics, diagnostics and autonomous systems.
Components
Capturing visual data from sensors: RGB cameras, depth cameras, 3D scanners or LiDAR.
Normalization, denoising, geometric correction and augmentation that prepare the image for analysis.
Computing features that describe the image — hand-engineered (SIFT, HOG) or learned by CNNs/ViTs.
The decision layer turning the representation into an output: a label, boxes, a segmentation mask or a depth map.
Implementation
A model trained on one image distribution degrades when deployment conditions (lighting, camera, environment) differ from the training data.
Unrepresentative data leads to systematic errors and unfair predictions for underrepresented groups or scenarios.
Variable lighting, shadows and partial occlusion of objects reduce detection and segmentation accuracy.
Evolution
Seymour Papert launches a student summer project to connect a camera to a computer and have it 'describe what it saw' — the symbolic beginning of the field.
Marr formulates a layered, computational model of vision, laying theoretical foundations for understanding visual perception.
The first efficient real-time face detector, based on Haar features and a cascade of classifiers.
A large-scale labeled dataset of millions of images that enabled training and benchmarking of deep vision models.
The deep convolutional network by Krizhevsky, Sutskever and Hinton dramatically lowers the classification error, launching the deep-learning era in computer vision.
Residual connections enable training of very deep networks, further raising image-recognition accuracy.
The transformer architecture is successfully applied directly to images split into patches, becoming an alternative to convolutional networks.
A model trained contrastively on image–text pairs enables zero-shot classification and opens the multimodal vision direction.
Hardware requirements
Deep-learning vision models (CNNs, ViTs) rely on massively parallel matrix multiplications and convolutions, which GPU tensor cores execute most efficiently.
Classical vision algorithms (filters, descriptors, geometric methods) also run on ordinary CPUs without acceleration.