Poor performance on large objects?

[hoonkai]

Is it expected that the detection of objects with a large surface (say above 60% of the image surface) is difficult to train with MobileNet? My experience is that the mAP and F1 are usually low if the training set contains mostly large objects.

Also, can MobileNet generalise from training on big objects/bounding boxes to detecting small objects/bounding boxes? Do the training objects need to be resized to match the test objects?

Any advice appreciated.