YOLOv3

YOLOv3 (You Only Look Once, v3) is a single-stage, state-of-the-art object detection architecture utilizing a Darknet-53 backbone for superior speed and accuracy in real-time applications.

YOLOv3, the third major iteration in the YOLO family, significantly improved detection performance, especially for small objects, over its predecessor (YOLOv2). The model uses the Darknet-53 feature extractor: a 53-layer residual network that boosts accuracy without sacrificing speed. Key to its design is the multi-scale prediction mechanism (detecting at three distinct scales: 13x13, 26x26, and 52x52 feature maps). Quantifiable results confirm its efficiency: on the COCO dataset, YOLOv3 achieved 28.2 mAP at 320x320 resolution, running in just 22 ms (demonstrating a speed advantage over models like SSD and RetinaNet).

https://arxiv.org/abs/1804.02767