search

HighwayWu / LASTED

Synthetic Image Detection
MIT License
51 stars 2 forks source link

Issue with Openset testset #3

Closed delyan-boychev closed 1 year ago

delyan-boychev commented 1 year ago

I have found that some of the synthetic images in Artist&Danbooru folders reproduce real photos, but not paintings as mentioned in the paper. Moreover, there is no annotation file in the repository. How can I use the Openset test set and where can I find the annotation file?

HighwayWu commented 1 year ago

Thanks for your asking. 1) Do you mean that some "painting" is classified as "photo" in the Artist&Danbooru folders? 2) Since the Openset test is just for testing (e.g., a binary classification/identification problem), the annotation is only one of "real" and "fake" as the file name related.

delyan-boychev commented 1 year ago

1) While reading the paper I understood that Artist&Danbooru is a set of paintings, but some of the synthetic images in Artist&Danbooru folders in Openset are synthetic photos, but not synthetic paintings as mentioned. 2) I'm trying to use it for multiclass classification - real photo, real painting, synthetic photo, synthetic painting.

HighwayWu commented 1 year ago
  1. Yes. The images in Artist&Danbooru should be all paintings. Please specific the name of the wrong image, and I'll double check.
  2. Sorry that the Openset is only for binary classification. Later the training dataset will be released, where multi-class evaluation could be conducted.
delyan-boychev commented 1 year ago

There are a lot of images, for instance: Openset_Testset/Artist_Danbooru_StyleGAN/synthetic_1064_StyleGAN.png

HighwayWu commented 1 year ago

Sorry for any confusion. In fact, the "Artist_Danbooru_StyleGAN/synthetic_1064_StyleGAN.png" means this is a synthetic image generated by StyleGAN, rather than a real image (photo/painting) from "ArtistDanbooru". Only the images with "real" prefix are from "Artist_Danbooru" (paintings) or "ImageNet_VISION" (photos).