IdolDanceNet:Indian Heritage idol Dance Pose Classification

S Kanimozhi, Sabari Nathan, Sasithradevi A; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 3463-3472

Capturing human movements and artistic expressions through sculptures and carvings has preserved cultural heritage for centuries. Bharatanatyam, one of India's most renowned classical dance forms, is rooted in 108 Karanas, which are extensively depicted in South Indian temple sculptures. However, these dance-specific sculptures remain largely undocumented in structured datasets, limiting their accessibility for computational analysis. To address this gap, we introduce IdolDance, a curated dataset featuring eight significant Bharatanatyam Karanas commonly found in temple architecture. We further propose IdolDanceNet, a deep learning model designed for robust classification of these sculptural postures. IdolDanceNet leverages depthwise convolution and attention mechanisms to extract spatial and contextual features while mitigating sculptural noise, pose variations, and texture inconsistencies. The model achieves 0.9917 precision and 0.9688 recall, demonstrating state-of-the-art performance in Bharatanatyam idol classification. Our framework facilitates digital heritage preservation, automated dance analysis, and pose retrieval systems, bridging the intersection of classical art and artificial intelligence.