To train a backbone for feature extractor, should I use padded pic or upscaled pic to train?

Question

As we know, conv input size is fixed. But when pre-training an object-detection backbone, one often uses a dataset that contains only target object