Seite - 114 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

both SIMD architectures have totally different instructions when it comes to data reordering. The most powerful data rearrange instructions of NEON are VTBL and VTBX. One the one hand, they are expensive in terms of execution cycles. On the other hand, several cases can be replaced by faster instructions. For example, the vector shift can be achieved using the NEON instruction VEXT and a register containing zero; broadcasting a single element can be done with VDUP. The newest ARM instruction set named ARMv8 should allow further performance optimizations by adding cross-lane instructions providing functionality exactly needed by discrepancy norm calculation like horizontal summation and taking minimum or maximum [24]. Though, it is impossible to estimate the achiev- able speedupwithoutpractical tests. 5. Conclusion WeanalyzedavariantofanimagesimilaritymeasurebasedonHermannWeyl’sdiscrepancyfromthe point of view of efficient implementation by exploiting redundancies in computing multiple integral images. Finally,weproposedan implementationbasedonvectorizationofprefixsumsandsummed- area tableswhichresults inaspeed-upfactor16compared toastandard integral imagebasedcompu- tation. Future research is left for checking parallelization and implementation optimizations also of othervariantsof Weyl’s discrepancy. References [1] B.A.Moser,“AsimilaritymeasureforimageandvolumetricdatabasedonHermannWeyl’sdis- crepancy.,” IEEETransactionsonPatternAnalysisandMachineIntelligence, vol.33,pp.2321– 9,nov2011. [2] B. Moser, G. Stu¨bl, and J.-L. Bouchot, “On a non-monotonicity effect of similarity measures.,” in SIMBAD (M. Pelillo and E. R. Hancock, eds.), vol. 7005 of Lecture Notes in Computer Sci- ence, pp.46–60, Springer, 2011. [3] H. Weyl, “U¨ber die Gleichverteilung von Zahlen mod. Eins,” Mathematische Annalen, vol. 77, pp.313–352, Sept1916. [4] B.Chazelle,Thediscrepancymethod: randomnessandcomplexity.CambridgeUniversityPress, 2000. [5] B.A.MoserandT.Natschlager, “Onstabilityofdistancemeasures forevent sequences induced bylevel-crossingsampling,”SignalProcessing, IEEETransactionson, vol.62,no.8,pp.1987– 1999,2014. [6] B.A.Moser,“Stabilityof threshold-basedsamplingasmetricproblem,”inEvent-basedControl, Communication,andSignalProcessing(EBCCSP),2015InternationalConferenceon, pp.1–8, IEEE,2015. [7] B. A. Moser, “The range of a simple random walk onZ: An elementary combinatorial ap- proach,” TheElectronicJournalof Combinatorics, vol. 21,no.4,pp.P4–10,2014. [8] J.-L.Bouchot,G.Stu¨bl, andB.Moser,“Atemplatematchingapproachbasedonthediscrepancy norm for defect detection on regularly textured surfaces,” in 10th International Conference on Quality Control by Artificial Vision, pp. 80000K–80000K, International Society for Optics and Photonics, 2011. 114