Long-term Labeling: Reusing Annotations for Changing Detections
Karsten Behrendt, Abhishek Mantha · 2019
Perception and sensor fusion systems yield intermediate 2D and 3D representations of detected objects. In combination with high-definition maps, automated driving systems use these detections to track, predict, and assign properties to all stationary and dynamic objects within a vehicle's sensor field of view. Perception and fusion systems are regularly updated to improve accuracy and computational efficiency. This often changes underlying properties of detected objects. Machine learning models trained on these characteristics will therefore deteriorate over time, as the current state of the system and the state of the system during annotation diverge over time. To solve this problem, we propose a novel solution to annotate groundtruth objects with target attributes in raw sensor data. This enables us to run our updated systems on stored sensor data, associate detections to groundtruth objects, and transfer target attributes. Annotated groundtruth data can be labeled once to train machine learning models for multiple iterations of a system, which decreases development costs and time spent on annotation efforts. As part of this work, we introduce comprehensive lists for spatial and temporospatial detection-to-groundtruth-object relations to highlight the complexity of adequately reusing groundtruth data for machine learning models and tracking systems. Finally, we describe a sample application for reusing annotations for a parked car classifier, achieving an accuracy of 95.0% after transferring parked car attributes to detections within our test set.