Slide the number of images you have. A few hundred is nowhere near enough to train 25 million weights from scratch, so you start from a trained model.
Almost nobody has ImageNet-scale data. Training a backbone from scratch is off the table for most real projects.
