Computer Vision

Computer vision is the field of AI concerned with getting a machine to extract meaning from images and video — not just storing or displaying pixels, but answering questions like "what object is this," "where is it in the frame," and "what is it doing over time." Like NLP, its modern era is dominated by one underlying technique: a CNN or vision transformer backbone extracts features from an image, and a task-specific head on top turns those features into whatever output the application actually needs. At a high level, that split looks like this:

Classification, detection, and segmentation are three different granularities of the same underlying question, and the right one depends entirely on what a downstream system actually needs to act on. Classification answers "what is the single dominant thing in this image" (used in simple quality-pass/fail checks); object detection draws a bounding box and label around every instance of interest (used anywhere a system needs to count or locate multiple objects, from retail inventory to autonomous-vehicle pedestrian detection); segmentation labels every individual pixel, giving an exact outline rather than a box (used anywhere precise shape matters, like medical tumor boundaries or a wafer defect's exact geometry). Picking a coarser task than the application needs wastes nothing but also tells a system less than it could know; picking a finer one than needed costs real extra compute and labeling effort for information nobody will use.

The granularity choice isn't just about accuracy — it directly sets labeling cost, which is often the real constraint on which task is even feasible. A classification label takes a human annotator seconds per image; a bounding box takes tens of seconds per object; a pixel-accurate segmentation mask can take minutes per image, especially for irregular shapes like biological tissue or manufacturing defects. This cost gradient is exactly why production systems often start with the cheapest task that solves the business problem, and only move to detection or segmentation once classification genuinely can't deliver the needed precision — not because finer-grained models don't exist, but because the labeled-data budget to train one well doesn't.

Computer vision backbones are trained once on a huge general dataset and then reused across wildly different downstream tasks, which is one of the field's biggest practical efficiency wins. A ResNet or ViT pretrained on ImageNet's 14 million general images learns broadly useful visual features (edges, textures, shapes) that transfer well to specialized domains with far less labeled data than training from scratch would need — a wafer-defect classifier or a medical-imaging model typically starts from an ImageNet-pretrained backbone and fine-tunes only on a comparatively small set of domain-specific labeled examples, which is often the difference between a project needing 500,000 labeled images or 5,000.

Read computer vision as one shared backbone pointed at increasingly expensive questions: classification, detection, and segmentation are the same visual features read out at three different costs and three different precisions, and the right choice for any given deployment is whichever granularity the labeling budget can actually afford to support at the accuracy the application needs — not automatically the most detailed one available.