Statistical vs. visual data generation in hand gesture recognition

摘要

A dataset with diverse training data is essence of the hand gesture recognition research. Most of the benchmarked datasets are limited in the number of signers and/or the number of each gesture try, which often result in over-fitting and poor generalization. Overcoming this challenge is often achieved by collecting a large number of exemplars for each hand gesture. This process is either expensive or impractical. Recently, synthetic data generation methods have been presented as a more reliable way to extend and enrich datasets. This paper proposes a comparative study of statistical and visual synthetic data generation. The visual synthetic data generation is executed by building synthetic 3D animated models using human figure design software. The experiments illustrate how the recognition accuracy will be changed when both methods used in enlarging the training data. The results show that in both cases recognition accuracy is enhanced, and in most cases, the visual synthetic data enlargement provides better improvement in the quality and diversity of the training data.