Comparison of the YOLOv3 and SSD Models Using a Balanced Dataset with Data Augmentation, for Object Recognition in Images

There are several models for object detection, among them the SSD and YOLO computer vision tools. These recognition systems are used to detect and classify objects in images or video frames in real time, with good performance. This article studies and compares the YOLOv3 and SSD MobileNet v2 algorithms for identifying objects in images. In order to achieve the intended objective, at first, the algorithms were trained and compared without data augmentation. After, the data augmentation was executed for improving the performance of the algorithms. Analyzing the results, a slightly better performance of the YOLOv3 model was observed, without performing data augmentation, although this model takes more time to complete the training for the same number of steps compared to the SSD MobileNet v2 model. On the other hand, when performing data augmentation, the SSD model is favored.