Comparison of 2D and 3D deep learning strategies for instance segmentation of wheat heads

One of the base parameters for wheat breeding is the wheat head count. In the context of high-throughput phenotyping, semantic and instance segmentation of single wheat head objects are important tasks to automatically derive wheat head parameters. A variety of approaches for such object extraction have been developed in the past years in 2D images and in 3D point clouds, especially in the context of deep learning. This paper compares three different strategies for high-quality detection with limited amount of reference data. The results point out that due to the higher variety of openly available 2D datasets, the transferability of 2D models to new datasets is better than for 3D models.