LookHere position encoding for ViTs. Code and pretrained models coming very soon.
ImageNet-HR dataset: https://huggingface.co/datasets/antofuller/ImageNet-HR
@misc{fuller2024lookhere,
title={LookHere: Vision Transformers with Directed Attention Generalize and Extrapolate},
author={Anthony Fuller and Daniel G. Kyrollos and Yousef Yassin and James R. Green},
year={2024},
eprint={2405.13985},
archivePrefix={arXiv},
primaryClass={cs.CV}
}