ETRI-Knowledge Sharing Plaform



논문 검색
구분 SCI
연도 ~ 키워드


학술대회 OpenCL-Darknet: An OpenCL Implementation for Object Detection
Cited 5 time in scopus Download 5 time Share share facebook twitter linkedin kakaostory
구용본, 유차영, 김성훈
International Workshop on Driving Computing Platform for Autonomous Vehicles (DrivComp) 2018, pp.631-634
17HS1800, 스마트카의 자율주행을 위한 실시간 센싱융합처리가 가능한 커넥티드 드라이빙 컴퓨팅 시스템 기술 개발, 김성훈
Object detection is a technology that deals with recognizing classes of objects and their location. It is used in many different areas, such as in face-detecting digital cameras, surveillance tools, or self-driving cars. These days, deep learning-based object detection approaches have achieved significantly better performance than the classic feature-based algorithms. Darknet [1] is a deep learning-based object detection framework, which is well known for its fast speed and simple structure. Unfortunately, like many other frameworks, Darknet only supports NVIDIA CUDA [2] for accelerating its calculations. For this reason, a user has only limited options for graphic card selection. OpenCL' (open computing language) [3] is an open standard for cross-platform, parallel programming of heterogeneous systems. It is available not only for CPUs, GPUs (graphics processing units), but also for DSPs (digital signal processors), FPGAs (field-programmable gate arrays) and other hardware accelerators. In this paper, we present the OpenCL-Darknet, which transforms the CUDA-based Darknet into an open standard OpenCL backend. Our goal was to implement a deep learning-based object detection framework that will be available for the general accelerator hardware and to achieve competitive performance compared to the original CUDA version. We evaluated the OpenCL-Darknet in AMD R7-integraged APU (accelerated processing unit) with OpenCL 2.0 and AMD Radeon RX560 with OpenCL 1.2 using a VOC 2007 dataset [4]. We also compared its performance with the original Darknet for NVIDIA GTX 1050 with CUDA 8.0 and cuDNN 6.0.
KSP 제안 키워드
Accelerated Processing Unit, CUDA 8.0, Competitive performance, Cross-Platform, Detection Approaches, Detection Framework, Digital camera, Digital signal processor(DSP), Fast speed, Feature-based, Field Programmable Gate Arrays(FPGA)