Hard Patches Mining for Masked Image Modeling
We observe that the reconstruction loss can naturally be the metric of the difficulty of the pre-training task.
Github: https://github.com/haochen-wang409/hpm
Paper: https://arxiv.org/abs/2304.05919v1
Dataset: https://paperswithcode.com/dataset/ade20k
https://t.me/DataScienceT