Computer Vision

Computer vision is the AI field that interprets images and video, turning pixels into structured facts the way people extract meaning from sight.

Tasks stack by detail. Image classification labels a whole image. Object detection finds and boxes many objects. Semantic segmentation labels every pixel by class. Instance segmentation separates individuals. Pose estimation finds body positions. Depth estimation rebuilds 3D structure from 2D frames.

Convolutional neural networks led the field for about a decade, learning edges, then textures, then parts, then objects, with translation-invariant filters. Vision Transformers now match or beat that by treating images as patch sequences and running attention, showing image-specific inductive biases are not strictly required at large scale.

CLIP and GPT-4V tie vision to language, so a model can understand an image from a text query with no extra labeled training. Products already depend on this: vehicles reading road scenes, medical tools spotting tumors, factory inspection, face and activity recognition, augmented reality, and crop monitoring from drones and satellites.

The field moved from lab demos to infrastructure inside many products.

Classification answers what is in the frame. Detection adds where. Segmentation adds which pixels. Pose adds joints. Depth adds distance. CNNs built those skills with local filters stacked into hierarchies. ViT showed patches plus attention can match that if you scale data. CLIP connects a photo to phrases without a labeled dataset for that task. GPT-4V answers questions about a picture.

The same stack now sits in cars, clinics, factories, cameras, AR headsets, and farm drones. That is infrastructure, not a demo reel. AlexNet (2012) cut ImageNet error enough that almost all production vision stacks became convolutional, then later Transformers.