80 likes | 88 Views
MPEG 112. Warsaw, July 2015. CDVS on mobile GPUs. Our Challenge. Compute CDVS descriptor from a stream video continuously Make pairwise matching between an entire database and the frame just acquired. Is it possible with an embedded system?. CDVS on mobile GPUs.
E N D
MPEG 112 Warsaw, July 2015 CDVS on mobile GPUs
Our Challenge • Compute CDVS descriptorfrom a stream video continuously • Makepairwisematchingbetweenanentire database and the frame just acquired Is it possible with an embedded system? CDVS on mobile GPUs
Feature Extraction in real time • CDVS Extraction from an image takes about two seconds on embedded systems such as smartphones or tablets • If we want to acquire information from a video stream this timing is inadequate • We need to use specific hardware Solution: GPU CDVS on mobile GPUs
CPU vs GPU GPUs have thousands of cores to process parallel workloads efficiently CDVS on mobile GPUs
GPU Programming Languages • Open ComputingLanguage • Itcontains the API to operate with the GPU • Very similar to CUDA language CDVS on mobile GPUs
MPEG CDVS Extraction pipeline Keypoint detection Feature selection Local descriptor computation Local descriptor compression Coordinate coding Local descriptors Image Global descriptor aggregation Global descriptor These three steps are implemented on GPU CDVS on mobile GPUs
TIM EYER • CDVS extraction using GPU hardware • Small database inside the smartphone • Pairwise matching operation between frame video and image stored into database CDVS on mobile GPUs