GDPR; object detection; dataset anonymization

Grade research has to be replicable, thus the used data need to be publicly available. Speaking, e.g., about object detection task, where image data for autonomous driving also contain privacy information such as faces and license plates, the publication of data may be harmful to captured people. The solution to the moral dilemma is to anonymize the data. In this study, our aim is to investigate the effect of various anonymization techniques on the performance of algorithms that use such data. We discuss anonymization methods that remove and replace privacy data and select three methods to replace the privacy data: blurring, permutation, and replacing the area with a constant value. We adopted the Cityscapes dataset from which we extracted areas containing privacy information and are the manner of the anonymization methods. Our benchmark involves three famous object detectors: YOLOv3, Mask R-CNN with ResNet-50 backbone, and Mask R-CNN with Swin-T backbone. The results show that the impact of anonymization methods on the performance is negligible and the impact is similar for both convolutional-based and transformer-based backbones.