Profile driven dataflow optimisation of mean shift visual tracking

Deepayan Bhowmik, Andrew Wallace, Robert Stewart, Xinyuan Qian, Greg Michaelson

Research output: Chapter in Book/Report/Conference proceedingConference contribution

1 Citation (Scopus)
101 Downloads (Pure)

Abstract

Profile guided optimisation is a common technique used by compilers and runtime systems to shorten execution runtimes and to optimise locality aware scheduling and memory access on heterogeneous hardware platforms. Some profiling tools trace the execution of low level code, whilst others are designed for abstract models of computation to provide rich domain-specific context in profiling reports. We have implemented mean shift, a computer vision tracking algorithm, in the RVC-CAL dataflow language and use both dynamic runtime and static dataflow profiling mechanisms to identify and eliminate bottlenecks in our naive initial version. We use these profiling reports to tune the CPU scheduler reducing runtime by 88%, and to optimise our dataflow implementation that reduces runtime by a further 43% — an overall runtime reduction of 93%. We also assess the portability of our mean shift optimisations by trading off CPU runtime against resource utilisation on FPGAs. Applying all dataflow optimisations reduces FPGA design space significantly, requiring fewer slice LUTs and less block memory.
Original languageEnglish
Title of host publication2014 IEEE Global Conference on Signal and Information Processing (GlobalSIP)
PublisherIEEE
ISBN (Electronic)9781479970889
DOIs
Publication statusPublished - 9 Feb 2015
Event2014 IEEE Global Conference on Signal and Information Processing - GA, Atlanta, United States
Duration: 3 Dec 20145 Dec 2014

Conference

Conference2014 IEEE Global Conference on Signal and Information Processing
Abbreviated titleGlobalSIP 2014
Country/TerritoryUnited States
CityAtlanta
Period3/12/145/12/14

Fingerprint

Dive into the research topics of 'Profile driven dataflow optimisation of mean shift visual tracking'. Together they form a unique fingerprint.

Cite this