[2022 Spring] ICCV 2021 Emerging Properties in Self-Supervised Vision Transformers (20213474)

[nooppi18]

좋은 리뷰 감사합니다.

몇 가지 추가하면 좋겠다고 생각한 것은

1. BYOL architecture 그림을 추가하면 선행연구를 이해하기 더 좋을 것 같습니다
2. 본 논문의 중요한 idea 가 collapsing 을 막는 centering 과 sharpening 인 것 같은데, 그 방법에 대해서 조금 더 자세히 설명해주시면 더 좋을 것 같습니다.

감사합니다

[enver1323]

Great review, but there are couple of moments, I would like to mention:

1. The top section should only contain a link to another language and not the whole block.
2. Possibly, it would be good idea to expand abbreviations, when using them the first time.
3. Typos might be corrected with the help of grammarly.com

[chrisahn99]

This was very interesting paper review on a very exciting and trending field of computer vision by showing Vision transformers trained using a novel self supervised training framework known as DINO.

I have not much comment on the overall architectures, except an unusual section that appeared in the middle of the paper: "ViT uses the transformer architecture to

Please introduce related work of this paper. Here, you need to list up or summarize strength and weakness of each work."

I don't know what this part is supposed to refer to, please check it and fix it. Thank you for your review!