Summary of the 2019 Activity Detection in Extended Videos Prize Challenge

Abstract

Despite previous data collection efforts and benchmark studies, progress in activity detection technologies has been slow, especially with applications that meet practical needs for the video analytics domain. In this paper, we discuss the results from the Activity detection in Extended Video Prize Challenge (ActEV-PC) that was sponsored by IARPA. The goal of the ActEV-PC was to promote robust automatic activity detection system development and to reduce the detection error rate. To examine the ability of activity detection systems in different aspects, we opened a competition to the public and ran evaluations (as a task under the ActivityNet workshop at CVPR 2019) with two different phases: an open leaderboard evaluation and a sequestered data evaluation. The Video and Image Retrieval and Analysis Tool (VIRAT) dataset was used for the open leaderboard evaluation while the Multiview Extended Video with Activities (MEVA) dataset was used for the sequestered data evaluation. Eighteen target activities were defined for detection. In this paper, we present results and findings from the two-phase ActEV-PC competition. Eighteen teams from academia and industry participated in the competitions and three top performers received a cash award (funded by IARPA). The winners were presented at the ActivityNet Workshop at CVPR 2019.