Let us cover three tasks on one example — a city-street photo: 1) CLASSIFICATION — "what is this image?": The model gives ONE label for the whole photo. E.g. "a city street" or, if the task is the main object, "a car". One question — one answer for the whole image. — The simplest. — Does not say WHERE the objects are and how many — only the general category. — Application: sorting photos by category, "defect/normal", scene-type detection, face mood. 2) OBJECT DETECTION — "what and where?": The model finds EACH object, draws a box (bounding box) around it and puts a label. On the street photo: a "car" box around each car, a "person" box around each pedestrian, a "sign" box around a sign. — Harder than classification: not one label but MANY objects, each with its position. — Gives the what, where and how many. — Application: counting objects (how many cars), video surveillance, car perception, retail shelf analytics. 3) SEGMENTATION — "which exact pixels belong to the object?": Even more precise than detection. Instead of a rectangular box — a PRECISE MASK: which specific pixels are the car, which are the road, which are the sky. Outlines the exact contour/shape, not a rough rectangle. — The most precise and complex of the three. — Needed when the exact shape/boundary matters: medical scans (the exact tumor contour), extracting an object for photo editing, precise road separation for cars. — Kinds: semantic (all "car" pixels marked as a car) and per-object (distinguishes individual cars). The key distinction in one phrase: — Classification: WHAT is in the picture (one label). — Detection: WHAT and WHERE (boxes around objects). — Segmentation: WHICH EXACT PIXELS (a shape mask). Growing precision = growing complexity = usually more data and computation needed. The practical meaning of the choice: take the task for the REAL need, no more complex than necessary. — Need to just categorize a photo? Classification — simpler, cheaper, less data. — Need to find and count objects, know their location? Detection. — Need the exact object shape/boundary? Segmentation — but it is more complex and demanding. A common mistake is taking a more complex task than needed (segmentation where classification would suffice): more data, computation, complexity without benefit. Determine which QUESTION you ask the image, and choose the minimally sufficient task type. There are other tasks too (OCR — text, face recognition, pose estimation, video tracking), but these three are the foundation of understanding CV. Understanding the difference, you correctly formulate the task and choose the tool.