TY - GEN
T1 - Reducto
T2 - 2020 Annual Conference of the ACM Special Interest Group on Data Communication on the Applications, Technologies, Architectures, and Protocols for Computer Communication, SIGCOMM 2020
AU - Li, Yuanqi
AU - Padmanabhan, Arthi
AU - Zhao, Pengzhan
AU - Wang, Yufei
AU - Xu, Guoqing Harry
AU - Netravali, Ravi
N1 - Funding Information:
Acknowledgements. We thank Frank Cangialosi, Amy Ousterhout, Anirudh Sivaraman, and Srinivas Narayana for their valuable feedback on earlier drafts of this paper. We also thank our shepherd, Ganesh Ananthanarayanan, and the anonymous reviewers for their constructive comments. This work is supported in part by NSF grants CNS-1613023, CNS-1703598, CNS-1943621, and CNS-1763172, and ONR grants N00014-16-1-2913 and N00014-18-1-2037.
Publisher Copyright:
© 2020 ACM.
PY - 2020/7/30
Y1 - 2020/7/30
N2 - To cope with the high resource (network and compute) demands of real-time video analytics pipelines, recent systems have relied on frame filtering. However, filtering has typically been done with neural networks running on edge/backend servers that are expensive to operate. This paper investigates on-camera filtering, which moves filtering to the beginning of the pipeline. Unfortunately, we find that commodity cameras have limited compute resources that only permit filtering via frame differencing based on low-level video features. Used incorrectly, such techniques can lead to unacceptable drops in query accuracy. To overcome this, we built Reducto, a system that dynamically adapts filtering decisions according to the time-varying correlation between feature type, filtering threshold, query accuracy, and video content. Experiments with a variety of videos and queries show that Reducto achieves significant (51-97% of frames) filtering benefits, while consistently meeting the desired accuracy.
AB - To cope with the high resource (network and compute) demands of real-time video analytics pipelines, recent systems have relied on frame filtering. However, filtering has typically been done with neural networks running on edge/backend servers that are expensive to operate. This paper investigates on-camera filtering, which moves filtering to the beginning of the pipeline. Unfortunately, we find that commodity cameras have limited compute resources that only permit filtering via frame differencing based on low-level video features. Used incorrectly, such techniques can lead to unacceptable drops in query accuracy. To overcome this, we built Reducto, a system that dynamically adapts filtering decisions according to the time-varying correlation between feature type, filtering threshold, query accuracy, and video content. Experiments with a variety of videos and queries show that Reducto achieves significant (51-97% of frames) filtering benefits, while consistently meeting the desired accuracy.
KW - deep neural networks
KW - object detection
KW - video analytics
UR - http://www.scopus.com/inward/record.url?scp=85094840598&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85094840598&partnerID=8YFLogxK
U2 - 10.1145/3387514.3405874
DO - 10.1145/3387514.3405874
M3 - Conference contribution
AN - SCOPUS:85094840598
T3 - SIGCOMM 2020 - Proceedings of the 2020 Annual Conference of the ACM Special Interest Group on Data Communication on the Applications, Technologies, Architectures, and Protocols for Computer Communication
SP - 359
EP - 376
BT - SIGCOMM 2020 - Proceedings of the 2020 Annual Conference of the ACM Special Interest Group on Data Communication on the Applications, Technologies, Architectures, and Protocols for Computer Communication
PB - Association for Computing Machinery
Y2 - 10 August 2020 through 14 August 2020
ER -