Effectiveness of Synthetic Image Data in Training Human Action Recognition Models

The growing trend of using large image datasets to support the training of computer vision algorithms in applications such as unmanned systems has seen a push towards the use of synthetic image data as an alternative source of data. Synthetic data has the potential enable the use of computer vision in applications that would previously have insufficient data to train a good model. However, the use of synthetic data to train machine learning models is not without caveats. The effectiveness of synthetic data as a training source, as compared with real data, is difficult to evaluate. Different papers that have explored the use of synthetic data have noted varying levels of effectiveness depending on the type of synthetic data tested and the type of model being trained. Outside of a general consensus that synthetic data is likely not detrimental, there is limited information available on what effect different synthesis parameters can have on the effectiveness of synthesised data. This paper evaluates the performance of composite synthetic data by training a human pose and action recognition model, investigating the effect different synthesis parameters have on a model trained using real and synthetic data.