A review: Comparison of performance metrics of pretrained models for object detection using the TensorFlow framework

Advances in parallel computing, GPU technology and deep learning facilitate the tools for processing complex images. The purpose of this research was focused on a review of the state of the art, related to the performance of pre-trained models for the detection of objects in order to make a comparison of these algorithms in terms of reliability, ac-curacy, time processed and Problems detected The consulted models are based on the Python programming language, the use of libraries based on TensorFlow, OpenCv and free image databases (Microsoft COCO and PAS-CAL VOC 2007/2012). These systems are not only focused on the recognition and classification of the objects in the images, but also on the location of the objects within it, drawing a bounding box around the appropriate way. For this research, different pre-trained models were re-viewed for the detection of objects such as R-CNN, R-FCN, SSD (single-shot multibox) and YOLO (You Only Look Once), with different extractors of characteristics such as VGG16, ResNet, Inception, MobileNet. As a result, it is not prudent to make direct and parallel analyzes between the different architecture and models, because each case has a particular solution for each problem, the purpose of this research is to generate an approximate notion of the experiments that have been carried out and conceive a starting point in the use that they are intended to give.