YOLO Net on iOS
YOLO Net on iOS
复制标题
iOS 上的 YOLO 网络
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
Maneesh Apte
中科院分区：
文献类型：
--
作者：
Maneesh Apte
我们的项目旨在研究在移动的设备上实现CNN进行真实的对象检测的速度和准确性之间的权衡。我们专注于修改YOLO架构，以实现移动的用例的最佳速度和准确性。具体来说，我们研究了SqueezeNet中使用的火层对分类性能的影响。虽然这一层确实提高了分类的速度，但准确性的下降太大，无法在移动的上使用。然后，我们使用Metal框架将性能最高的tiny-YOLO net部署到iOS应用程序中，该网络的运行速度为8-11 FPS，mAP为51%。该应用程序可以从VOC数据集中检测20个类别，并将类别真实的翻译成5种不同的语言。这种实时检测本质上是一个移动的Rosetta Stone，它在本地运行，因此与WiFi无关。
Our project aims to investigate the trade-offs between speed and accuracy of implementing CNNs for real time object detection on mobile devices. We focused on modifying the YOLO architecture to achieve optimal speed and accuracy for mobile use cases. Specifically, we investigated the effect of a fire layer, as used in SqueezeNet, on classification performance. While this layer did increase the speed of classification, the decrease in accuracy was too great for use on mobile. We then deployed the highest performing tiny-YOLO net, which operates at 8-11 FPS with a mAP of 51%, to an iOS app using the Metal framework. The app can detect the 20 classes from the VOC dataset and translate the categories into 5 different languages in real time. This real-time detection is essentially a mobile Rosetta Stone that runs natively and is thus wifi independent.