Estimating efficiency in the categorization of biological motion

Abstract

It is often claimed that point light displays provide sufficient information to easily recognize properties of the actor and action being performed. We examined this topic by obtaining estimates of human efficiency in the categorization of movement. We began by collecting a database of human arm movements from 13 males and 13 females that contained multiple repetitions of knocking, waving and lifting movements done both in an angry and a neutral style. For each movement, 3D position data was recorded at 6 locations on the arm and head at a rate of 60 Hz. Point light displays of each individual for all of the six different combinations were presented to participants who were asked to judge the gender of the model. Results of human performance were compared to the output of automatic pattern classifiers based on artificial neural networks designed and trained to perform the same classification task on the same movements. A value of d-prime was obtained for both the judgments by human participants and the neural networks. Estimates of the upper bound of efficiency were defined as the ratio of d-prime squared for the human and neural network. Results for gender recognition indicate efficiencies on the order of a few percent and vary with the movement condition and the movement representation input to the neural network.