Crop or original image better for training?

[nathnim]

As an example, I have a large image contains multi objects like traffic vehicles. The first way I annotate all objects in the original image. The second way I crop the large image in to small images which one contain one object and mark for single object only.

For the first way I have to mark manually. For the second way I have a semi-automatic tool to do so it is save time.

But the problem I wandering is that which way is better for quality of the training model? If no much different I will go with the second way to mark objects for saving time.

Thanks in advance!