Does larger object perform better in segmentation?

According to the fcn model I implemented and the pascal voc benchmark hereI find that object which have larger size in image receive better iou score or ap score in testset. Can anyone give me some tips about why this happened?

Small location errors for large objects will have less influence on the IOU than on small objects.
Here is a good explanation on the IOU score.

Additionally, I think small objects might generally be “hard” examples in comparison to large ones.
Maybe Focal loss (RetinaNet) might help in your use case.

Thank you ptrblck , Always can I find useful information in your answer!