Multimedia Systems (14321882)31(2)
In recent years, action recognition has witnessed significant advancements. However, most existing approaches heavily depend on the availability of large amounts of video data, which can be computationally expensive and time-consuming to process especially in real-time applications with limited computational resources. Utilizing too few frames instead, may lead to the loss of crucial information. Therefore, selecting a few frames in a way that preserves essential information poses a challenge. To address this issue, this paper proposes a novel video clip embedding technique called Hybrid Embedding. This technique combines the advantages of uniform frame sampling and tubelet embedding to enhance recognition with few frames. By employing a transformer-based architecture, the approach captures both spatial and temporal information from limited video frames. Furthermore, a keyframe extraction method is introduced to select more informative and diverse frames, which is crucial when only a few frames are available. In addition, the region of interest (ROI) in each RGB frame is cropped using skeletal data to enhance spatial attention. The study also explores the impact of the number of frames, different modalities, various transformer models, and the effect of pretraining in few-frame human action recognition. Experimental results demonstrate the effectiveness of the proposed embedding technique in few-frame action recognition. These findings contribute to addressing the challenge of action recognition with limited frames and shed light on the potential of transformers in this domain. © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2025.
