This paper aims to guide practitioners in selecting a detection architecture that balances speed, memory usage, and accuracy for specific applications and platforms. It investigates various trade-offs in modern convolutional object detection systems, focusing on the "meta-architectures" Faster R-CNN, R-FCN, and SSD. These architectures are implemented in Tensorflow to trace the speed/accuracy trade-off curve by varying feature extractors and other parameters like image size. The paper highlights that reducing the number of proposals in Faster R-CNN can significantly improve speed without substantial loss in accuracy. SSDs are found to be less sensitive to the quality of the feature extractor compared to Faster R-CNN and R-FCN. The study also identifies optimal points on the trade-off curve where gains in accuracy come at the cost of speed. Novel combinations of meta-architectures and feature extractors are reported, including Inception networks in Faster R-CNN and MobileNet in SSD, which were used to train the winning entry of the 2016 COCO object detection challenge. The paper concludes with practical insights and techniques to improve speed without sacrificing accuracy.This paper aims to guide practitioners in selecting a detection architecture that balances speed, memory usage, and accuracy for specific applications and platforms. It investigates various trade-offs in modern convolutional object detection systems, focusing on the "meta-architectures" Faster R-CNN, R-FCN, and SSD. These architectures are implemented in Tensorflow to trace the speed/accuracy trade-off curve by varying feature extractors and other parameters like image size. The paper highlights that reducing the number of proposals in Faster R-CNN can significantly improve speed without substantial loss in accuracy. SSDs are found to be less sensitive to the quality of the feature extractor compared to Faster R-CNN and R-FCN. The study also identifies optimal points on the trade-off curve where gains in accuracy come at the cost of speed. Novel combinations of meta-architectures and feature extractors are reported, including Inception networks in Faster R-CNN and MobileNet in SSD, which were used to train the winning entry of the 2016 COCO object detection challenge. The paper concludes with practical insights and techniques to improve speed without sacrificing accuracy.