Impact of Anonymization on Vehicle Detector Performance
Lukas Schnabel, Stephan Matzka, Martin Stellmacher, Michael Pätzold, Elmar Matthes · 2019
Due to strict data protection regulations in the European Union, many companies face great challenges regarding the gathering, processing and storage of image data for the development of driving functions based on convolutional neural networks (CNN). To meet this challenge, de-identification can be employed to remove personal information. This paper assesses the impact of anonymization on the detection of vehicles. A dataset is compiled and anonymized with different types of de-identification. A CNN object detector is trained on the original and anonymized versions of the dataset. Evaluation shows that the impact of anonymization depends on variousfactors, including type and size of the object, de-identification method and share of the object type in the dataset. In addition, we show that anonymization with pixelization has a rather insignificant impact on the performance of the object detector. Therefore, it can be considered a feasible approach to tackle the challenge of gathering image data on the open road while complying to current data protection regulations.