While vision transformers are able to solve a wide variety of computer vision tasks, no pre-training method has yet demonstrated the same scaling laws as observed in language models. Autoregressive models show promising results, but are commonly trained on images that are cropped or transformed into square images, which distorts or destroys information present in the input. To overcome this limitation, we propose NARAIM, a vision model pre-trained with an autoregressive objective that uses images in their native aspect ratio. By maintaining the native aspect ratio, we preserve the original spatial context, thereby enhancing the model's ability to interpret visual information. In our experiments, we show that maintaining the aspect ratio improves performance on a downstream classification task.
@article{arxiv.2410.10012,
title = {NARAIM: Native Aspect Ratio Autoregressive Image Models},
author = {Daniel Gallo Fernández and Robert van der Klis and Răzvan-Andrei Matişan and Janusz Partyka and Efstratios Gavves and Samuele Papa and Phillip Lippe},
journal= {arXiv preprint arXiv:2410.10012},
year = {2024}
}
Comments
Accepted to NeurIPS, see https://openreview.net/forum?id=7Iuh8VWU66