YOLO (You only look once) revolutionized the computer vision arena. The first version of YOLO was released in 2016 by Joseph Redmon et. al and it smashed benchmarks both in terms of speed and accuracy. When it comes to object detection, YOLO has been a favorite of Data Scientists and Machine Learning engineers and the go-to model when it comes to segmenting entities in images.
Since it was launched, YOLO had many new iterations that improved several setbacks of previous versions, namely:
- Improved architecture of the underlying deep learning models.
- Implemented alternatives to improve performance, such as data augmentation techniques.
- Migrated the original YOLO code to use pytorch training and deployment frameworks.
- Improved detection mechanisms of small objects.
The last version of YOLO is YOLO v9 (https://arxiv.org/abs/2402.13616). one important thing to be aware is that every computer vision and object detection model is evaluated on two parameters: Accuracy (defined by metrics related to computer vision segmentation) and Speed (defined by latency in the inference). One example of how CV algorithms are evaluated is shown below: