Two Recognition Models for Thai Dancing Data Set

Human activities recognition has been developed using various techniques for optimal recognition and many data sets have also been developed for this task. However, some Thai culture activities cannot be recognized by the techniques developed by oversea data sets. That is the main motivation for collecting a new data set about Thai dancing in this work. We compare the performance between 2 recognition techniques, i.e. 3D CNN and LSTM against this task. Five classes from 16 volunteers are collected in this data set. Then, two existing models from each technique are trained by the Thai dancing data set with 16 batch size and 50 epochs using 7:3 to represent the ratio between the number of training data and test data. The final results show that LSTM model achieved 0.1250 accuracy which is better than 3D CNN obtaining the accuracy score of 0.0800.