Evaluation of Approaches for Action Similarity Measurement

Programming by Demonstration has been recently proposed as a way for a robot learning tasks from human demonstrations where action recognition is a crucial step in the procedure. Based on this concept a model-free approach for object manipulation was proposed by Aksoy et al.[1]. In specific the approach classifies actions by observing object-interaction changes based on video segmentation. However the segmentation suffers from various difficulties such as motion blur complex environment over- and under- segmentation. For this reason we simulate and evaluate the Aksoy et al.''s method. Additionally we adapt a kernel based representation into Aksoy et al.''s method. The experiments shows the new method improves action recognition rate significantly.