The Impact of Relative Positional Embeddings

Hey, first off, thanks for sharing this work! I noticed that you're actively using Relative Positional Embeddings (RPE) for all attention operations [here](https://github.com/hustvl/ViTMatte/blob/main/configs/common/model.py#L34) and [here](https://github.com/hustvl/ViTMatte/blob/main/modeling/backbone/vit.py#L75). As the RPE are intended to emulate the convolutional layer shift-invariance property. I wonder what was the impact of this choice? Looking at your arxiv paper, I haven't seen any explicit mention of this choice or an ablation study. If you could share why RPE was important and if it is needed, I'd appretiate, thanks!