Variable Scale Gesture Recognition: A Dataset and Comprehensive Analysis

Dynamic gesture recognition plays an important role in natural human-computer interaction. Gesture image sequences can express more complex interaction intentions. Currently, the majority of dynamic gesture recognition algorithms are trained using fixed-length gesture sequences. However, the rate, amplitude, and angle of gestures differ across individuals, which presents a challenge to the effectiveness of existing dynamic gesture recognition algorithms. To address this challenge, we propose a variable scale gesture dataset. This dataset contains six types of dynamic gestures with different sequence lengths. In the experiments, we utilized major algorithms, including Two-Stream, C3D, CRNN and I3D, to test and evaluate their performance. Although current algorithms have been able to achieve high recognition rates for fixed-length sequences, they are unable to directly recognize variable-length sequences. The dataset includes four types of static gestures that can be used for dynamic gesture keyframe recognition and start/stop frame determination. The dataset proposed in this paper is expected to advance research in variable scale gesture recognition.