[2022 spring] CVPR 2021 DatasetGAN: Efficient Labeled Data Factory with Minimal Human Effort (20226042)

[stk14570]

기존 StyleGAN 구조를 활용하여 synthetic data에 대한 label generation을 수행한다는 흥미로운 내용의 논문이었습니다. 리뷰 format에는 큰 문제가 없어보입니다. 리뷰 감사드립니다.

[Alex-C137E]

Interesting paper about a relevant issue. The text is great and the illustration make it pleasant to read. I don't really have any comment otherwise. Good job !

[Victor02091]

Very interesting paper with pertinent images and formulas that makes it very enjoyable to read.

If I understood well these artificially generated labeled images will after be used to test/train other AI algorithm. Then how do we know that artificial images will perform without any bias on these other algorithms ? how do we know that their use should be the same as real images ?

[chrisahn99]

The generalization performance can be verified in the results section: the authors have shown that by using this synthetic dataset to train segmentation networks, they have better generalization capacity (thus low bias) compared to other fully supervised (on human labeled datasets) networks.