The Effects of Different Image Parameters on Human Action Recognition Models Trained on Real and Synthetic Image Data

Keith Man, Javaan S. Chahl, Wolfgang Mayer, Asanka G. Perera

Abstract: The growing use of computer vision in many fields has seen an increasing need for visual data, both real and synthetic, to support the training and testing of computer vision models. To this end, the need for high quality datasets has also increased. That said, the quality of image data for the purpose of computer vision training remains a comparatively unexplored field. There is little information on how image parameters may affect data quality, whether such parameters affect real and synthetic data equally, and how such parameters may interact with each other. The goal of this study was to test the effect of basic image data parameters on real and synthetic data and the resultant effect such parameters have on training a human action recognition model. A human action dataset with two distinct actions was selected and individual image parameters were modified to produce a total of 247 data subsets. Testing a total of 247 human action recognition models, trained on the data subsets, it was found that image parameters affect real and synthetic data differently. The impact of parameters, such as resolution and colour space, varies not just on whether the image data is real or synthesised, but also on the human action being performed. Hybrid datasets created by mixing real and synthetic data, up to 70% synthetic data, are able to produce models with similar performance levels to models trained on real data, without the negative effects of some parameters that could be observed in synthetic only datasets. Lastly, the mislabelling of training data has a large impact on action prediction accuracy, even at low percentages of mislabelled data.