Instead of being driven by downstream tasks, our model called MaeFuse utilizes a pretrained encoder from Masked Autoencoders (MAE), which facilities the omni features extraction for low-level ...