Application of CUDA and Tensor Kernels in Object Detection and Recognition Problems

The purpose of research is to study a way to accelerate the processes of object recognition in images using hardware video accelerators, which include kernels CUDA and TENSOR.

Methods. The method of software interaction with hardware computing means of a video accelerator with CUDA and RT cores using the Python programming language was used; Docker containers from NVIDIA GPU Cloud (NGC), OpenCV to run the feed from the camera and TensorRT to speed up the output of the data stream, the single shot detection network with InceptionV2 as the backbone. To investigate the performance and accelerate the recognition processes, an environment is designed that interacts with equipment and a number of libraries for rendering images. Within this environment, object recognition processes take place, using various computing cores CUDA and TENSOR, as well as various algorithms and accuracy classes.

Results. The use of TENSOR kernels accelerates the recognition process when using FP16 accuracy, as well as when using the combined FP16 and FP32 accuracy. Using single precision INT8 shows significantly better performance when using TensorRT on an accelerator with TENSOR cores.

Conclusion. The capabilities of hardware video accelerators have been investigated when they are used for object recognition tasks. A study of the performance of the CUDA and TENSOR kernels when interacting with the TensorRT engine has been carried out. A software solution for the interaction of video accelerators with CUDA and TENSOR cores with the TensorRT engine is proposed. This solution demonstrates high rates of object recognition speed with INT8 accuracy and slightly lower performance when using FP32 accuracy.

1. Informatsionnye tekhnologii i vychislitel'nye sistemy. Vychislitel'nye sistemy. Komp'yuternaya grafika. Raspoznavanie obrazov. Matematicheskoe modelirovanie [Information technology and computing systems. Computing systems. Computer graphics. Pattern recognition. Mathematical modeling]; ed. by S. V. Emelyanov. Moscow, Lenand Publ., 2015.

2. p.

3. Matematicheskie modeli sotsial'no-ekonomicheskikh protsessov. Modelirovanie kharakteristik deyatel'nosti otraslevykh i regional'nykh podsistem. Dinamicheskie sistemy. Matematicheskie problemy dinamiki neodnorodnykh sistem. Informatsionnye tekhnologii. Trudy ISA RAN [Proceedings of ISA RAS. Mathematical models of socio-economic processes. Modeling the performance characteristics of sectoral and regional subsystems. Dynamical systems. Mathematical problems of the dynamics of inhomogeneous systems. Information technologies]; ed. by S. V. Emelyanov. Moscow, Lenand Publ., 2015. 112 p.

4. Shmaglit L. A., Golubev M. N., Ganin A. N., Khryashchev V. V. [Gender classification by facial image]. Tsifrovaya obrabotka signalov i ee primenenie (DSPA-2012). Doklady 14 Mezhdunarodnoi konferentsii [Digital signal processing and its application. Reports of the XIV International Conference (DSPA-2012)]. Moscow, 2012, vol. 2, p. 425. (In Russ.)

5. Eliseeva I. I., Rukavishnikov V. O. Gruppirovka, korrelyatsiya, raspoznavanie obrazov (statisticheskie metody klassifikatsii i izmereniya svyazei) [Grouping, correlation, pattern recognition (statistical methods of classification and measurement of connections)]. Moscow, RGGU Publ., 2014. 144 p.

6. Boyko I. A., Guryanov R. A. Raspoznavanie ob"ektov na osnove videosignala, poluchennogo s kamery, ustanovlennoi na podvizhnoi platforme [Object recognition based on the video signal received from a camera installed on a mobile platform]. Molodoi uchenyi = Young Scientist, 2013, no. 6 (53), рр. 34-36.

7. Rabotyagov A. Osnovy raspoznavaniya obrazov [Fundamentals of pattern recognition]. Moscow, LAP Lambert Academic Publishing, 2014. 273 p.

8. Fukunaga K. Vvedenie v statisticheskuyu teoriyu raspoznavaniya obrazov [Introduction to the statistical theory of pattern recognition]. Moscow, Nauka Publ., 2013. 368 p.

9. Sistemnoe modelirovanie. Naukometriya i upravlenie naukoi. Raspoznavanie obrazov. Т-udy ISA RAN [Proceedings of ISA RAS: System modeling. Scientometrics and Science Management. Pattern Recognition]; ed. by S. V. Emelyanov. Moscow, Lenand Publ., 2015. 104 p.

10. Dudarev V. A. Metody raspoznavaniya obrazov v komp'yuternom konstruirovanii neorganicheskikh soedinenii [Methods of pattern recognition in computer design of inorganic compounds. Dudarev]. Moscow, Synergy Publ., 2014. 325 p.

11. Emelyanov S. V. Informatsionnye tekhnologii i vychislitel'nye sistemy. Vychislitel'nye sistemy. Komp'yuternaya grafika. Raspoznavanie obrazov. Matematicheskoe modelirovanie [Information technology and computing systems. Computing systems. Computer graphics. Pattern recognition. Mathematical modeling]; ed. by S. V. Emelyanov. Moscow, Mir Publ., 2015, is. 2. 662 p.

12. Milenky A. V. Klassifikatsiya signalov v usloviyakh neopredelennosti. Statisticheskie metody samoobucheniya v raspoznavanii obrazov [Classification of signals in conditions of uncertainty. Statistical methods of self-learning in pattern recognition]. Moscow, Soviet radio Publ., 2017. 328 p.

13. Tsvetkov A. A., Shorokh D. K., Zubareva M. G., eds. [Object recognition algorithms]. Tekhnicheskie nauki: problemy i perspektivy. Materialy IV Mezhdunarodnoi nauchnoi konferentsii (g. Sankt-Peterburg, iyul' 2016 g.) [Technical sciences: problems and prospects. Materials of the IV International Scientific Conference (St. Petersburg, July 2016)]. St. Petersburg, Own Publishing House, 2016, pp. 20-28. (In Russ.)

14. Avtomaticheskoe upravlenie i vychislitel'naya tekhnika. Raspoznavanie obrazov [Automatic control and computer technology. Pattern recognition]. Moscow, Mechanical Engineering Publ., 2016, is. 10. 256 p.

15. Informatsionnye tekhnologii i vychislitel'nye sistemy. Vychislitel'nye sistemy. Matematicheskoe modelirovanie. Raspoznavanie obrazov. Prikladnye aspekty informatiki [Information technology and computing systems. Computing systems. Mathematical modeling. Pattern recognition. Applied Aspects of Informatics]; ed. by S. V. Emelyanov. Moscow, Higher school Publ., 2014, is. 3. 160 p.

16. Grenander W. Lektsii po teorii obrazov [Lectures on the theory of images]. Moscow, Mir Publ., 2014, vol. 1. 571 p.

17. Obnaruzhenie, raspoznavanie i opredelenie parametrov obrazov ob"ektov. Metody i algoritmy [Detection, recognition and determination of the parameters of object images. Methods and algorithms]; ed. by A. V. Korennov. Moscow, Radiotekhnika Publ., 2012. 112 p.

18. Potapov A. Avtomaticheskii analiz izobrazhenii i raspoznavanie obrazov [Automatic image analysis and pattern recognition]. Moscow, LAP Lambert Academic Publishing, 2017. 292 p.

19. Grenander W. Lektsii po teorii obrazov [Lectures on the theory of images]. Moscow, Mir Publ., 2016, vol. 2. 342 p.

20. Grenander W. Lektsii po teorii obrazov [Lectures on the theory of images]. Moscow, Mir Publ., 2012, vol. 3. 432 p.

21. Informatsionnye tekhnologii i vychislitel'nye sistemy. Obrabotka informatsii i analiz dannykh. Programmnaya inzheneriya. Matematicheskoe modelirovanie. Prikladnye aspekty informatiki [Information technology and computing systems. Information processing and data analysis. Software engineering. Mathematical modeling. Applied Aspects of Informatics]; ed. by S. V. Emelyanov. Moscow, Lenand Publ., 2015. 104 p.