Segmenting objects with indistinct edges, with application to aerial imagery of vegetation

Image segmentation mask creation relies on objects having distinct edges. While this may be true for the objects seen in many image segmentation challenges, it is less so when approaching tasks such as segmentation of vegetation in aerial imagery. Such datasets contain indistinct edges, or areas of mixed information at edges, which introduces a level of annotator subjectivity at edge pixels. Existing loss functions apply equal learning ability to both these pixels of low and high annotation confidence. In this paper, we propose a weight map based loss function that takes into account low confidence in the annotation at edges of objects by down-weighting the contribution of these pixels to the overall loss. We examine different weight map designs to find the most optimal one when applied to a dataset of aerial imagery of vegetation, with the task of segmenting a particular genus of shrub from other land cover types. When compared to inverse class frequency weighted binary cross-entropy loss, we found that using weight map based loss produced a better performing model than binary cross-entropy loss, improving F1 score by 4%.