Object Detection in AI & Artificial Intelligence

While simple classification output is a single label, Object Detection outputs a Bounding Box $[x, y, w, h]$ for every detected instance. This process, known as Localization, requires the model to not only recognize features but also understand their spatial boundaries within the image grid. This is fundamental for tasks where the quantity and position of objects are critical, such as inventory counting or autonomous navigation.

Early detection methods (like R-CNN) were slow because they used a multi-stage pipeline: first proposing regions and then classifying them. YOLO (You Only Look Once) and SSD (Single Shot Detector) revolutionized the field by performing both steps in a single forward pass through the network. By treating detection as a regression problem over a fixed grid, these models achieve frame rates high enough for real-time video processing on standard hardware.

To evaluate how well a box fits an object, we use Intersection over Union (IoU), which measures the overlap between the prediction and the ground truth. Because models often predict multiple boxes for the same object, we apply Non-Maximum Suppression (NMS). NMS identifies the box with the highest confidence and 'suppresses' all nearby boxes that have a high IoU with it, ensuring each object is labeled exactly once.