Keeping track of the kids: a deep dive into object detector fairness for pedestrians of different ages

As autonomous driving gains increasing prevalence, it is important to ensure these vehicles are safe for everyone who comes across them. While overall performance is a key aspect, that is not the whole picture. Fairness is another crucial topic that needs to be measured when evaluating computer vision models. In our paper, we highlight an important area in which similar groups are not treated equally by current state-of-the-art object and pedestrian detectors. By grouping pedestrian data according to bounding box size and looking for patterns across metrics, we find that children are successfully detected less frequently than adults and that the reason for this disparity is not simply due to their smaller stature, as hypothesized. Additionally, we find that object detectors perform better than pedestrian detectors but are less fair. To this end, we suggest that more work is needed to identify the underlying causes of this complex disparity, and we offer a few possible avenues for doing so.