Portal de Periódicos da CAPES

Classification of Indian Classical Dance 3D Point Cloud Data Using Geometric Deep Learning

2021; Springer Nature; Linguagem: Inglês

10.1007/978-981-33-6862-0_7

2194-5357

Ashwini Naik, M. Supriya,

Human Motion and Animation

Indian classical dances have many unique postures that require to be identified and classified correctly. Though many classification techniques exist for two-dimensional dance images, there is a need to classify the three-dimensional images as it is still on the evolving side. Geometric deep learning is one of the growing fields in machine learning and deep learning. It enables to learn from complex type of data represented in the form of graphs and 3D objects (manifolds). Deep learning algorithms like convolution neural networks (CNN) and recurrent neural networks (RNN) have achieved higher performance on the broad range of problems. Using these algorithms, one can also classify the images. Deep learning algorithm works well for Euclidean data such as points, lines, and planes. CNN cannot be implemented on the non-Euclidean data such as graphs and 3D object (manifolds), and thus, neural network architecture that can learn from non-Euclidean data is required. In the proposed work, implementation of geometric deep learning is done on 3D image data represented as point cloud. PointNet architecture will work efficiently with point cloud data. This architecture has been used to classify Indian classical dance point cloud data into five dance forms, namely Bharatanatyam, Odissi, Kathak, Kathakali, and Yakshagana.