Classification, Detection, Segmentation

Tasks

Part of: Computer Vision Basics

"Is there a cat in this photo?" "Where is the cat?" "Which exact pixels are cat?" Those are three different questions, and in computer vision they are three different tasks with three different outputs. Pick the wrong one and you either get less than you need or pay for far more than you use. This lesson is the map of what vision systems can be asked to produce. What it is There is a clean ladder of vision tasks, each one finer-grained than the last: - Classification answers " what is this image? " with one label for the whole picture. Input: an image. Output: a class like cat (often with a confidence). It does not say where anything is. - Detection answers " what objects are here and where? " with a bounding box around each object plus its label. Input: an image. Output: a list of (label, box) pairs. The box is four numbers, x1, y1, x2, y2. - Segmentation answers " which pixels belong to what? " with a per-pixel mask . Input: an image. Output: a label for every single pixel. It traces the exact outline, not just a rectangle. Each step up gives more spatial precision and costs more to produce. How it works The same backbone (the CNN or vision layers from earlier lessons) feeds all

Challenge: The Detection Report