The 2nd YouTube-8M Large-Scale Video Understanding Challenge

Joonseok Lee, Apostol Natsev, Walter Reade, Rahul Sukthankar, George Toderici

Abstract: We hosted the 2nd YouTube-8M Large-Scale Video Understanding Kaggle Challenge and Workshop at ECCV’18, with the task of classifying videos from frame-level and video-level audio-visual features. In this year’s challenge, we restricted the final model size to 1 GB or less, encouraging participants to explore representation learning or better architecture, instead of heavy ensembles of multiple models. In this paper, we briefly introduce the YouTube-8M dataset and challenge task, followed by participants statistics and result analysis. We summarize proposed ideas by participants, including architectures, temporal aggregation methods, ensembling and distillation, data augmentation, and more.