Date : 2017.10.02(Mon) 05:00 P.M.
Locate : EB5. 533
Presenter : Trunghai Do
Title : YOLO implementation on ARM Mali GPU
Author : Trunghai Do, Jemin Lee, Hyungshin Kim
Nowadays, convolutional neural networks become the heart of many computer vision solutions to solve a wide range of tasks including image classification, object detection and segmentation. In this paper, we present YOLO implementation on ARM Mali-T628 MP6 GPU of ODROID-XU4. Original YOLO algorithm was implemented by Darknet framework dedicated for NVIDIA GPU. Unlike Darknet, we use CK-Caffe as a deep learning backend to run YOLO on ARM-based GPUs. Additionally, we replace all fully connected layers which act as detection layers in YOLO with convolutional layers to reduce the model size. We train our GoogLeNet-based and Fast YOLO-based models on canonical PASCAL VOC2007 + VOC2012 "trainval" (training + validation) set on NVIDIA GTX-1080 and experiment on PASCAL VOC2007 test set on ODROID-XU4. As a result, GoogLeNet takes 7.5 seconds per image and Fast YOLO takes 1.6 seconds per image at inference time.
Article source: http://eslab.cnu.ac.kr/en/Mobile/101-YOLO-implementation-on-ARM-Mali-GPU.html