VANT at TRECVID 2018

We propose a system for activity detection, which utilizes the Action Tubelet (ACT) Detector to localize activities in video data. Our network is trained for all of activities in the ActEV dataset with a backbone convolutional neural network pre-trained on the ImageNet dataset. We inserted a thresholding module to the original ACT framework to adapt detector to the ActEV task, since activities in this task appear more sparsely distributed than those in the action detection task. Our result was 0.882 in mean-p miss@0.15rfa at the AD Leaderboard Evaluation.