Towards memory-efficient inference in edge video analytics