{Wanli Ouyang Yali Wang Zhiyong Wang Ding Liang Lei Bai Luping Zhou Zhipeng Yu Yu Guo Peiqin Zhuang}

Abstract

Motion modeling is crucial in modern actionrecognition methods. As motion dynamics like moving temposand action amplitude may vary a lot in different video clips,it poses great challenge on adaptively covering proper motioninformation. To address this issue, we introduce a MotionDiversification and Selection (MoDS) module to generatediversified spatio-temporal motion features and then select thesuitable motion representation dynamically for categorizing theinput video. To be specific, we first propose a spatio-temporalmotion generation (StMG) module to construct a bank ofdiversified motion features with varying spatial neighborhoodand time range. Then, a dynamic motion selection (DMS)module is leveraged to choose the most discriminative motionfeature both spatially and temporally from the feature bank.As a result, our proposed method can make full use ofthe diversified spatio-temporal motion information, whilemaintaining computational efficiency at the inference stage.Extensive experiments on five widely-used benchmarks,demonstrate the effectiveness of the method and we achievestate-of-the-art performance on Something-Something V1 & V2that are of large motion variation