Question: Landmarking Model

[imaccormick275]

Hi

I recently read your paper here

https://www.groundai.com/project/facial-landmark-detection-for-manga-images/#bib.bib10

I was wondering if the landmarking model you trained is available anywhere?

Iain

[oaugereau]

The facial landmark labels for the Manga109 dataset can be downloaded from this Github.

For more detail about the content of this dataset, please refer to the paper: https://arxiv.org/abs/1811.03214

We made only the labels available. To have access to Manga109 images, the permission can be asked here: http://www.manga109.org/en/

This project has been done in Osaka Prefecture University. The responsible person can be contacted here: http://kyoindb.osakafu-u.ac.jp/html/100453_en.html

[imaccormick275]

Some quick questions about the annotations provided:

• What is the best way to match your landmarking coordinates to a given face in Manga109? I see in Manga109 there is a unique face id - but not in your json file.
• Each json starts with userxx_ - does this relate the user that annotate the file? In the case of duplicate e.g. "MeteoSanStrikeDesu_021" what is the best way to handle such a case? Take the average of all users who annotated or just arbitrarily select one?
• The annotations provided are for artists from M-R (MeteoSanStrikeDesu - RisingGirl.) I just wanted to check that this file is complete - will I be able to extract 1446 images mentioned in the paper from the jsons provided?

[oaugereau]

Manga 109 changed the format of their dataset recently. The new Manga 109 dataset include face annotations: http://www.manga109.org/en/annotations.html