Performance Evaluation Between Tiny Yolov3 and MobileNet SSDv1 for Object Detection

Abstract

Object detection plays a crucial role in the field of computer vision. It is viewed as a challenging task as it identifies instances of objects from a particular class in digital images or videos. However, since the invention of deep learning methods, the performance of object detection has significantly improved. They are now able to learn semantic, high-level, and deeper features to address existing issues found in traditional architectures. In this paper, an evaluation framework has been proposed to assess the performance of Tiny Yolov3 and MobileNet SSD v1 for detecting people. In addition, both Tiny Yolov3 and MobileNet SSD v1 consist of a lightweight architecture that eliminates the expensive computation to run the models in real time detection using a NON-GPU platform. A fair comparison was made between the pre-trained models by using the two available datasets which are COCO and PASCAL VOC. The model’s performance was evaluated in a classroom scenario, where people were detected and counted. A mobile application was built to view the detection results and its performance was assessed when used with deep learning models. To have a more expansive evaluation, different parameters such as platform, cameras, and conditions were considered. From those parameters, different test cases were formulated and tested to determine which models excel the most and where. Following the evaluation, this paper proposes an evaluation framework for MobileNet SSD v1 and Tiny Yolov3 and provides a domain recommendation for future applications.