Fig. 2

Mask R-CNN training and validation losses during training for 200 epochs on ImgOld and ImgNew maize cob images from the Peruvian genebank. a Loss curves leading to the best model M104 in epoch 95. c With the same scaling on the y-axis, parameter combination M109 shows substantial overfitting as indicated by much higher validation losses resulting in an inferior model based on \(AP@\)[.5:.95]. b Loss curves of M104 with a zoomed scale on the y-axis, highlighting the mask loss as highest contributor to overall loss, indicating that masks are most difficult to optimize. Other losses, like class loss or bounding box loss, are of minor importance