Datasets & Annotations
Recording, processing and annotating the right data is half the work in developing new machine learning applications. SafeAD utilizes visual SLAM, 3D reconstruction, deep neural networks and custom annotation tools to facilitate the annotation process.
To make full use of a dataset the recording vehicle has to provide precisely time stamped data as well as intrinsically and extrinsically calibrated sensors. SafeAD uses a dedicated hardware trigger and custom calibration tools for best results.
Recording multiple cameras and LiDARs for hours requires a large amount of data storage. SafeAD uses highly efficient data compression to reduce the required storage by a factor of 30. Further, we anonymize faces and license plates on the fly and encrypt the data using AES-256 for GDPR compliance.
SafeAD reduces the manual annotation effort to a minimum by employing a CNN-based autolabeling pipeline in combination with an accausal tracker. A smart keyframe selection algorithm in our annotation tool suggests which keyframes the human annotator should label for best results. To increase the awareness and to facilitate the annotation of static parts of the scene we load a 3D reconstructed model of the environment into the annotation tool.