Making automated driving safer: RELiABEL project launched
Making automated driving safer: RELiABEL project launched
This work provides a sandbox environment that allows rapid prototyping of active learning methods for object detection, saving up to a factor of 32 in computational time. This is complemented by useful evaluation metrics, baselines for comparison, and baseline results on PascalVOC and BDD100k that can be reused by the community.
Preprint: https://arxiv.org/pdf/2212.10836
During our work on [1], we realized that label quality plays a particularly important role in situations where only small amounts of training data are available. We combine AI-guided label checking with simple uncertainty-guided data selection and show that
1) label checking significantly improves model performance (as a function of labeling effort) and
2) uncertainty-guided data selection and label checking are indeed synergistic.
Preprint: https://arxiv.org/pdf/2310.00372
Both papers were conducted as a collaboration between the Technical University of Berlin and the University of Wuppertal.