This paper presents a comprehensive analysis of the speed/accuracy trade-offs in modern convolutional object detectors. The goal is to provide a guide for selecting an architecture that achieves the right balance of speed, memory, and accuracy for a given application and platform. We investigate various ways to trade accuracy for speed and memory in modern object detection systems. We present a unified implementation of Faster R-CNN, R-FCN, and SSD, which we view as "meta-architectures," and trace the speed/accuracy trade-off curve by using alternative feature extractors and varying critical parameters such as image size. On one extreme, we present a detector that achieves real-time speeds and can be deployed on a mobile device. On the opposite end, we present a detector that achieves state-of-the-art performance on the COCO detection task. We explore the speed/accuracy trade-off of modern detection systems in an exhaustive and fair way. We focus on single-model/single-pass detectors, which do not use ensembling, multi-crop methods, or other "tricks." We focus on test-time performance rather than training time. We find that using fewer proposals for Faster R-CNN can significantly speed it up without a big loss in accuracy, making it competitive with SSD and R-FCN. We also find that SSD performance is less sensitive to the quality of the feature extractor than Faster R-CNN and R-FCN. We identify sweet spots on the accuracy/speed trade-off curve where gains in accuracy are only possible by sacrificing speed. We evaluate various feature extractors, including VGG-16, Resnet-101, Inception v2, Inception v3, Inception Resnet (v2), and MobileNet. We find that using fewer proposals can significantly reduce running time without a big loss in accuracy. We also find that using a lower input resolution can reduce inference time but may lower accuracy. We find that the performance of SSD is less reliant on its feature extractor's classification accuracy compared to Faster R-CNN and R-FCN. We also find that the performance of detectors is highly correlated with their classification accuracy. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find thatThis paper presents a comprehensive analysis of the speed/accuracy trade-offs in modern convolutional object detectors. The goal is to provide a guide for selecting an architecture that achieves the right balance of speed, memory, and accuracy for a given application and platform. We investigate various ways to trade accuracy for speed and memory in modern object detection systems. We present a unified implementation of Faster R-CNN, R-FCN, and SSD, which we view as "meta-architectures," and trace the speed/accuracy trade-off curve by using alternative feature extractors and varying critical parameters such as image size. On one extreme, we present a detector that achieves real-time speeds and can be deployed on a mobile device. On the opposite end, we present a detector that achieves state-of-the-art performance on the COCO detection task. We explore the speed/accuracy trade-off of modern detection systems in an exhaustive and fair way. We focus on single-model/single-pass detectors, which do not use ensembling, multi-crop methods, or other "tricks." We focus on test-time performance rather than training time. We find that using fewer proposals for Faster R-CNN can significantly speed it up without a big loss in accuracy, making it competitive with SSD and R-FCN. We also find that SSD performance is less sensitive to the quality of the feature extractor than Faster R-CNN and R-FCN. We identify sweet spots on the accuracy/speed trade-off curve where gains in accuracy are only possible by sacrificing speed. We evaluate various feature extractors, including VGG-16, Resnet-101, Inception v2, Inception v3, Inception Resnet (v2), and MobileNet. We find that using fewer proposals can significantly reduce running time without a big loss in accuracy. We also find that using a lower input resolution can reduce inference time but may lower accuracy. We find that the performance of SSD is less reliant on its feature extractor's classification accuracy compared to Faster R-CNN and R-FCN. We also find that the performance of detectors is highly correlated with their classification accuracy. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that the performance of detectors on small objects is less reliable than on large objects. We find that