An Approach for Fusing Two Training-Datasets with Partially Overlapping Classes

Joshua Niemeijer, Federico Battistella, Gurucharan Srinivas, Andreas Leich

Abstract: Supervised deep learning techniques in image processing require training data, typically consisting of manually labeled ground truth annotations. Since manual labeling is costly, using as many existing training datasets as possible is necessary. This paper introduces a novel approach for combining training datasets into a new one. The naive approach to this is a plain concatenation of the existing datasets. However, this approach fails with partially overlapping datasets when certain annotated instances specific to one dataset also appear in the other dataset -without their annotations. Therefore, we present a novel method for combining existing training datasets using a pseudo-labeling technique with uncertainty quantification. The effectiveness of our method is evaluated by fusing two datasets consisting of partially overlapping traffic sign annotations in street view images. The results demonstrate that the pseudo-labeling errors weigh less than those introduced by the naive fusion. Furthermore, our work provides evidence for practitioners to use a pseudo-labeling-based fusion technique with uncertainty quantification rather than naively combining training datasets into a new one.