An Empirical Study of Object Detectors and Its Verification on the Embedded Object Detection Model Competition

As one of the most active research fields in computer vision, object detection has been boosted by recent advancements of deep learning. However, performance improvements are often accompany with more resource consumption. There has been a rising interest in building a detector that runs well on embedded systems, i.e., with a better tradeoff among accuracy, speed, and resource consumed. To this end, we analyze the pipeline of popular object detectors, with an emphasis on taking typical evaluation metrics from the above three aspects to evaluate both the selection of backbone network and the effects of input resolution. As a result, we propose an efficient model by combining MobileNet and SSD. The model is verified on the embedded object detection model competition, where we gain the best accuracy on the scooter detection.