We implement YOLO, a state of the art real-time object detection system. The goal is to regress the bounding boxes and class probabilities with a single network. The ground truths are defined based on grid cells that split up the image. Each ground truth is assigned to a ground truth bounding box with the channels defined as: P(objectness), x, y, w, h, and class probabilities (pedestrian, traffic light, car). Then each image is fed through the model architecture and a loss function is used for optimization. Finally, in inference, we perform postprocessing using non maximum suppression.
