Performance Analysis of YOLOv3, YOLOv4 and MobileNet SSD for Real Time Object Detection

Abstract

Real time identification of specific objects by machine learning approach has exhibited excellent outcomes; however, in the actual databases real-life images are prone to in-focus, unnecessary, blurred, noisy, shaky and jittery images. These problems create a great deal of wiggle room in success evaluation of real-time object recognition algorithms. To address these issues, object recognition mainly comprises applying the technique of You Only Look Once (YOLO). The unique feature of YOLO is the possibility of identifying all objects in the image. Other algorithms, unfortunately, are incapable of coping with the complexity of the whole images. Fast object detection is a key feature of YOLO classifiers since they are able to locate the best boxes and provide probabilities for each box in a very short time, surpassing the speed of other algorithms. YOLO localizes objects on pictures with its high level of precision. This study provides the real-time performance analysis of YOLOv3, YOLOv4 and MobileNet SSD for object detection. In a recent experiment, different object detection models were compared for their accuracy and speed. YOLOv4 boasted the highest precision (99.50%) and recall (98%) but took the longest to process an image (1.91 seconds). YOLOv3 offered a good balance, achieving a mean average precision (mAP) of 96.5% and recall of 94.6% with a faster inference time of 0.345 seconds. MobileNet SSD prioritized speed, delivering the quickest inference time (0.145 seconds) but with the lowest accuracy (precision: 91.4%, recall: 88.6%). This highlights the trade-off between accuracy and speed in object detection. YOLOv4 prioritizes accuracy, YOLOv3 offers a balance, and MobileNet SSD prioritizes speed.