YOLOv3

YOLOv3 is the third version of the You Only Look Once (YOLO) family of real-time object detection systems. Introduced in 2018 by Joseph Redmon and Ali Farhadi, it improves upon its predecessors by increasing accuracy while maintaining high speed. The model is implemented in the Darknet framework and uses Darknet-53 as its backbone, a 53-layer convolutional network that provides robust feature extraction for detection tasks.

Architecture and predictions: YOLOv3 employs a feature pyramid network approach to generate predictions at three different scales, corresponding to feature maps of varying resolutions. After processing the input through Darknet-53, the network upsamples and combines features to detect smaller objects on finer maps while still predicting larger objects on coarser maps. At each scale, it predicts bounding boxes using three anchor boxes per location. Each predicted box includes an objectness score and class probabilities. Unlike some detectors that use a single softmax over classes, YOLOv3 applies independent sigmoid activations for each class, enabling multi-label predictions.

Anchors and training: Anchors are determined via k-means clustering on the training data, and the model is trained on the COCO dataset, which comprises 80 object classes. The training optimizes both the objectness score and the class probabilities, with non-maximum suppression applied during inference to produce final detections.

Performance and impact: YOLOv3 is designed to balance speed and accuracy for real-time applications, delivering fast detections on standard GPUs. It quickly became a widely adopted baseline in both research and industry, influencing subsequent YOLO iterations and contributing to the broader development of efficient real-time object detectors.