Computer Vision

Computer vision is the field concerned with extracting meaning from images and video — deciding what is in a picture, where it is, what changed between frames, and what text or structure it contains. The task family is worth naming precisely, because vendors blur them and they differ enormously in difficulty. Classification says what an image is. Object detection says what is in it and draws boxes. Segmentation labels every pixel. Optical character recognition pulls out text. Pose estimation locates joints. Tracking follows an object across frames. A product need stated as "we want the AI to look at the photo" usually resolves to one of those, and identifying which one changes the data you must collect and the accuracy you can expect. Deep convolutional networks made the field practical in the 2010s by learning visual features instead of relying on hand-designed ones, and multimodal models have since collapsed a lot of it into prompting: you can now hand an image to a general model and ask a question in words, which is transformative for one-off and long-tail tasks. For high-volume, latency-sensitive, or precision-critical work, a purpose-trained detector still wins on cost, speed, and consistency — the same trade-off that governs language tasks. For SaaS builders the common applications are unglamorous and valuable: reading a receipt or an ID document, checking that an uploaded photo meets a listing standard, moderating user-generated images, extracting fields from a scanned form, verifying a product is present in a marketplace photo. Each of those is a pipeline rather than a model call, and the pipeline is where the work lives — handling rotation and glare and phone-camera variation, deciding the confidence threshold below which a human looks at it, and logging what the model saw so a wrong decision can be explained later. Two failure modes deserve planning for. Vision models are sensitive to distribution shift in ways that are hard to anticipate: lighting, camera model, resolution, and background all move the accuracy you measured. And they carry real privacy and fairness weight, because the inputs are pictures of people and the errors are not evenly distributed across them.