Oberseminar Bildverarbeitung, Computersehen und Computergraphik
|Prof. Dr.-Ing. Thomas Brox, Prof. Dr.-Ing Matthias Teschner, apl. Prof. Dr. Olaf Ronneberger
Geb. 52, SR 02-017
Herr Lukas Halbritter
Institut für Informatik, Universität Freiburg
berichtet über die Ergebnisse seiner Bachelorarbeit:
Evaluation of dataset characteristics for deep network training
Current state-of-the-art deep networks for disparity estimation are capable of generating quite accurate disparity maps from stereo images. The amount of computergenerated datasets used for training these neural networks is steadily increasing. Besides the advantages that come with this kind of data like dense and precise ground truth there are also problems that arise from using these datasets. The computergenerated data differs from real-world images in many ways and describes not the exact domain as images shot with a real camera. Artificial data does not suffer from lense distortion or a low dynamic range in the first place, unless it is added during the rendering process. In this work, we modify synthetic training data in a way that it matches a real-world test dataset better. We train neural networks on data modified in various ways and compare the results to those of networks trained on unmodified data.