The effects of model certainty, test-time augmentation, and their trade-offs on leaf segmentation and counting

Plant phenotyping tasks such as leaf segmentation and counting are fundamental to the study of phenotypic traits. Deep supervised learning has been prevalent in recent works proposing better performing models for these tasks. Despite good efforts, one of the main challenges for presenting better methods is the limitation of labelled data availability. Such limitation turned the field's main efforts to techniques for augmenting existing data sets, leaving some aspects of model training under-discussed. This paper explores the effect of these training aspects and present simple methods that led to a top-1 ranking model in both the Leaf Segmentation Challenge and the Komatsuna public data sets. The model has competitive performance while been arguably simpler than others recently proposed. The experiments brought insights related to the positive effects of model cardinality and test-time augmentation on the segmentation of single-class objects with high occlusion, and the data distribution of these data sets.