Alternating wavelet channel and spatial attention mechanism for online video-based Indian classical dance recognition

Recognizing poses in Indian classical dance (ICD) from live performance videos is a challenging task due to variable conditions such as lighting, scaling, costumes, and capture angles. This paper introduces the alternating wavelet channel and spatial attention (AWCSA) model, designed to enhance generalization capabilities of convolutional neural network (CNN) features by integrating low and high-frequency wavelet information through alternating channel and spatial attention mechanisms. The AWCSA model effectively captures both structural and textural cues from video frames, leading to improved dance pose classification. Experimental results show that the AWCSA model outperforms existing methods on the Bharatanatyam Onstage Online Indian Classical Dance Video Dataset (BOOICDVD23) and the ‘Let’s Dance’ dataset.