Modeling Extremely Large Images with $x$T

Handling large images is a challenge due to the loss of information when down-sampling or cropping. The xT framework introduces a nested tokenization approach to divide images into regions and sub-regions for better processing. It combines region encoders and context encoders to extract features at different scales and provide a comprehensive understanding of the image. The framework allows for modeling extremely large images and achieves great results in tasks such as species classification.