YOLO is a single-stage object detector: instead of proposing regions and then classifying them, it predicts object locations and labels in one forward pass. This makes it a practical choice when detection speed matters, including real-time applications. section 12.2.2 section 12.4
A grid-based prediction
Given a 448 × 448 image, a YOLO head can divide the image into a 6 × 6 grid. Each grid location predicts one bounding box, a confidence score, and class probabilities. In the demonstrated setup, the model predicts 91 classes, producing separate box and class outputs for every grid cell. section 12.2.2
A practical YOLO-style model consists of:
- A ConvNet backbone that extracts high-level visual features from the image. section 12.2.2
- A confidence value trained to be high where an object exists and near zero in empty grid locations. section 12.2.2
The demonstrated implementation uses a pretrained ResNet backbone and matching KerasHub preprocessing, then resizes images to 448 × 448 before attaching the YOLO prediction head. Training data comes from COCO, with bounding boxes normalized for the resized images. section 12.2.2
