Using synthetic data to train convolutional neural networks for the case of hand detection

Barbara Aljaž (2018) Using synthetic data to train convolutional neural networks for the case of hand detection. EngD thesis.

Preview

PDF Download (3115Kb)

Abstract

Convolutional neural networks require a large amount of data for training that need to be collected and annotated. Methods used to enlarge learning dataset usually include different augmentations, but in this thesis we researched the possibility of using artificially generated data samples. We created them using a three dimensional model and automatically controlled parameters that influenced captured images. We worked on the example of human hand detection and evaluated our detector on two datasets of real images for a touch-less interface human-computer interaction scenario. We compared it with a detector trained on real life data and analyzed the differences. Results of the experiment are promising and present many opportunities for further development of such training technique.

Actions (login required)