Evaluating the Real-World Application Efficacy of MobileNet Models

This experimental study explores the abilities of MobileNet and its three variants within the sphere of object classification for object detection under different lighting. Our research trains every model on the ‘Car Object Detection’ dataset with adjustments to lighting, weather conditions, and urban or rural settings, which represent real-life situations more accurately. We outline specific alterations made to architecture and methods used during training that were meant to increase adaptability across different environments while maintaining accuracy, too. As a result, this work achieved remarkable results, and our best-performing algorithm attained a 97% validation accuracy rating according to tests carried out under various environmental conditions. Through lightweight convolutional networks for object detection, it becomes clear that such type was not only effective but also resource efficient, hence applicable in dynamic settings requiring real-time operation with limited resources.