ViTamin: Designing Scalable Vision Models in the Vision-Language Era

arXiv:2404.02132v2 Announce Type: replace
Abstract: Recent breakthroughs in vision-language models (VLMs) start a new page in the vision community. The VLMs provide stronger and more generalizable feature embeddings compared to those from ImageNet-pretrained models, thanks to the training on the large-scale Internet image-text pairs. However, despite the amazing achievement from the VLMs, vanilla Vision Transformers (ViTs) remain the default choice for the image encoder. Although pure transformer proves its effectiveness in the text encoding area, it remains questionable whether it …

abstract achievement arxiv community cs.cv designing embeddings feature however image imagenet internet language language models page pretrained models scalable scale text training type vision vision-language models vision models vlms