Copy and mix instance object to different image so that model can learn instance features with varying background
With copy & paste augmentation, large scale jittering also help models to learn more various features
Pretraining could hurt accuracy with strong enough data augmentation
Kaiming He, Ross Girshick, and Piotr Dollar. Rethinking imagenet pre-training. In ICCV, 2019
Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin D. Cubuk, and Quoc V. Le. Rethinking pre-training and self-training. In NeurIPS, 2020
If data augmentation of fine-tuning stage can provide better information than the pretrained information, pretrained information could mislead model from learning rich information brought by data augmentation
However, many of these experiment is conducted on large scale models such as ResNet-50 FPN backbone. The result could be different for mobile architectures such as MNasNet, MobileNet
