YOLOv3 is an incremental improvement over previous versions, featuring design changes and a more accurate network. It is faster than SSD and comparable in accuracy. YOLOv3 achieves 57.9 AP50 on a Titan X in 51 ms, outperforming RetinaNet by 3.8×. The model uses a new feature extractor, Darknet-53, which is more powerful than Darknet-19 but more efficient than ResNet-101 or ResNet-152. It predicts bounding boxes at three scales, using a feature pyramid network approach. YOLOv3 also uses multilabel classification for object detection, which better models overlapping labels. The model performs well on the old detection metric of AP50 but struggles with higher IOU thresholds. YOLOv3 is faster and better than other models on the AP50 metric. The authors tried various techniques but found that their current formulation is effective. They also fixed a data loading bug in YOLOv2, improving performance by 2 mAP. The paper discusses the importance of considering the ethical implications of computer vision research. YOLOv3 is a good detector, fast, and accurate, though it is not as strong as other models on the COCO average AP metric. The authors argue that the COCO metric may not reflect real-world performance and suggest alternative metrics. The paper concludes with a call for responsible use of computer vision technology.YOLOv3 is an incremental improvement over previous versions, featuring design changes and a more accurate network. It is faster than SSD and comparable in accuracy. YOLOv3 achieves 57.9 AP50 on a Titan X in 51 ms, outperforming RetinaNet by 3.8×. The model uses a new feature extractor, Darknet-53, which is more powerful than Darknet-19 but more efficient than ResNet-101 or ResNet-152. It predicts bounding boxes at three scales, using a feature pyramid network approach. YOLOv3 also uses multilabel classification for object detection, which better models overlapping labels. The model performs well on the old detection metric of AP50 but struggles with higher IOU thresholds. YOLOv3 is faster and better than other models on the AP50 metric. The authors tried various techniques but found that their current formulation is effective. They also fixed a data loading bug in YOLOv2, improving performance by 2 mAP. The paper discusses the importance of considering the ethical implications of computer vision research. YOLOv3 is a good detector, fast, and accurate, though it is not as strong as other models on the COCO average AP metric. The authors argue that the COCO metric may not reflect real-world performance and suggest alternative metrics. The paper concludes with a call for responsible use of computer vision technology.