Mobile 241 Views

by Do Trung Hai on 2018-05-29 15:09:39

Date : 2018. 05. 30 (Wed) 13:00

Locate : EB5. 533

Presenter : Trunghai Do


Nowadays, convolutional neural networks (CNNs) become the center of many computer vision solutions to solve a variety of tasks. However, memory and computation are two of the most important characteristics of deep neural networks. These characteristics make neural networks difficult to effectively deploy on limited hardware resources such as embedded systems. Furthermore, to deploy models for devices and update them regularly, the model size needs to be small.

In this thesis, we propose DroidDet a small fully convolutional neural network. Our DroidDet adopts You Only Look Once (YOLO) object detection algorithm for the ARM Mali-T628 MP6 GPU of ODROID-XU4. In order to build DroidDet, we not only replace all the fully connected layers that act as detection layers in YOLO with convolutional layers but also rearrange some of the convolutional layers to reduce the model size and keep the computational cost as small as possible. Compared to SqueezeDet, a 10.5MB object detection model takes 3.829 seconds for inferencing an image, and compared to Tiny YOLO, another 101.8MB model takes 1.651 seconds for each image, our DroidDet is as small as SqueezeDet and as fast as Tiny YOLO. Namely, the DroidDet is 14.9MB and takes 1.722 seconds for each image.

In addition, the YOLO algorithm was originally developed by using the Darknet framework for NVIDIA GPUs. Unlike Darknet, we use CK-Caffe as a deep learning backend to deploy YOLO on ARM-based GPUs. We have trained our models on the canonical PASCAL VOC2007 and VOC2012 “trainval” (including training and validation) set on NVIDIA GTX-1080 GPU. Then we experimented on PASCAL VOC2007 released test set on ODROID-XU4. Moreover, we extend our approach to apply for object detection on real-time video streams and measure processing rate using FPS metric.

Article source:


File Attachment: Download 1