Activity Detection in Untrimmed Videos Using Chunk-based Classifiers

Abstract

Activity detection in untrimmed videos - the process of detecting and localizing human activities in potentially long videos - is a challenging problem in computer vision. We propose an algorithm which is based on the proposition that despite the differences between activity classification and detection, a strong classifier can still be used to achieve state-of-the-art performance in detection by breaking the video into multiple overlapping chunks and classifying each individually. We further introduce two new auxiliary tasks which we call chunk inclusion and localization. The outputs of these tasks, when carefully applied, can be used to dramatically improve performance. We call our method Chunk Aggregation. It is straight-forward to implement and use, and is agnostic to the backbone activity classification architecture used. We also demonstrate the effectiveness of chunk association by presenting results and a series of ablation experiments on the THUMOS’14 and ActEV datasets.