Frugal Event Data: How Small Is Too Small? a Human Performance Assessment with Shrinking Data

Abstract

When designing embedded computer vision systems with limited computational budget, one often needs to take care of the size of input data. In recent years, however, event cameras have shown increasingly large sensor sizes. How small can event data be, while preserving sufficient information for the task at hand? We present in this paper a study to assess and compare human performance in a gesture classification task using event data. Original event data from IBM’s DVS128 Gesture dataset is downscaled with several spatial and temporal methods, and the classification performance on 4 classes is measured with human participants. The contributions of this paper are 3-fold: (1) we establish a size threshold under which the human performance falls behind the chance level, (2) we compare several spatial and temporal event downscaling methods and show that all methods give unequal data quality, and (3) we highlight some unexpected discrepancies in a comparison between human vs machine performance. To the best of our knowledge, this is the first human perception study with event data.