Real-time Detection of Activities in Untrimmed Videos

Joshua Gleason, Carlos Domingo Castillo, Rama Chellappa

Abstract: Real-time detection of spatio-temporal sparse activities in untrimmed videos is a challenging problem. In this work, we present the details of our proposed solution. We begin with a slow baseline implementation of a previously state-of-the-art system [13] and redesign it to achieve real-time performance for detecting 37 activities in the ActEV19 Sequestered Data Leaderboard [4]. This is primarily achieved by introducing speed related hyperparameters into the baseline approach. A tradeoff analysis is performed to assist in hyperparameter selection which results in a real-time, high quality action detection system. Our system achieves an AUDC score of 0.476 on the ActEV19 Sequestered Data Leaderboard.