AUTHORS: Worawat Choensawat, Woong Choi, Hiroyuki Sekiguchi and Kozaburo Hachimura

ABSTRACT: Segmentation is a key process of the motion-capture-data retrieval systems, which have been widely used in media technologies. This paper introduces a new method to improve the segmentation accuracy of complex motion data such as dancing. The proposed method is a neural network which is achieved as non-linear classifier in many applications. Use of SMOTE technique can overcome a defective case of neural networks in case the data set is seriously imbalanced. An imbalanced dataset is where the instances of one class far outnumber the other class. In our case, the ratio of segmentation and non-segmentation classes equal to 1:10. Our experimental results show the increasing percentage of both precision

and recall comparing between using the neural network with and without SMOTE technique. The proposed method also provides higher accuracy than the previous works up to 50%.