Vision Transformers Need Registers - Fixing a Bug in DINOv2?
In this video we explain the research paper titled Vision Transformers Need Registers by Meta AI, which was written by authors that were part of DINOv2 paper. This paper describes a phenomenon in DINOv2 outputs that does not happen in DINOv1, which is referred as artifacts.
This phenomenon was discovered to be relevant for more large foundational computer vision models in addition to DINOv2, which are OpenCLIP and DeiT.
We start with essential background about visual features, and then explain what are the artifacts, what is their impact and when do they appear. Afterwards, we'll describe the solution suggested in the paper to avoid these artifacts, using register tokens in vision transformers, and then we'll see how this method performs.
👍 Please like & subscribe if you enjoy this content
Blog post - https://aipapersacademy.com/vision-tr...
Paper page - https://arxiv.org/abs/2309.16588
DINOv2 video summary & full review - https://aipapersacademy.com/dinov2-fr...
-----------------------------------------------------------------------------------------------
Support us - https://paypal.me/aipapersacademy
We use VideoScribe to edit our videos - https://tidd.ly/44TZEiX (affiliate)
We use ChatPDF to analyze research papers - https://www.chatpdf.com/?via=ai-papers (affiliate)
-----------------------------------------------------------------------------------------------
Chapters:
0:00 Agenda
0:53 Background
2:21 Artifacts
6:31 ViT Registers
7:40 Results
8:38 Conclusion