Multi-class hair image database with Ground Truth: Figaro 1K
A project done in collaborator with Umar Riaz Muhammad, PhD (link).
Before you download the data, please note: The pictures in the dataset were collected from the web for the purpose of carrying out not-for-profit scientific experiments and are not University of Brescia property. Any use of the dataset, other than ‘fair use‘, must be negotiated with the pictures’ owners. University of Brescia is not responsible for the content nor the meaning of the images.
In order to carry out hair analysis in the wild, a database with unconstrained view images containing various hair textures is needed. The scarcity of open and available databases pushed us to build and make publicly available Figaro1k, an extension of Figaro, an annotated novel multi-class (straight, wavy, curly, kinky, braids, dreadlocks and short; each containing 150 images) image database for hair analysis in the wild.
Being images different in size and aspect ratio, a normalisation procedure has then been applied: the employed normalisation factor is chosen so as to reduce the size of the maximum squared area inscribed in the hair region to 227×227 pixels.
Patch-F1k is an auxiliary database used only for training hair detection at patch level. It is provided together with Figaro1k and contains 1,050 pure hair texture images and 1,050 pure non-hair texture images, for a total of 2,100 images of size 227×227. Both hair and non-hair images are divided in the same way as Figaro1k, i.e. 840 for training and 210 for testing. In our experiments we use only the training set. In order to make the two datasets compatible, hair patches in the training set of Patch-F1k are directly extracted from 840 images of Figaro1k training set (first row of Figure 5), while non-hair patches are partially extracted from Figaro1k (420 images), and partially from VOC2012 (again 420), for a total of 840 non-hair images.
The 7 classes are distributed in this order:
- straight: frame00001-00150
- wavy: frame00151-00300
- curly: frame00301-00450
- kinky: frame00451-00600
- braids: frame00601-00750
- dreadlocks: frame00751-00900
- short-men: frame00091-01050
References
Muhammad, U. R., Svanera, M., Leonardi, R., & Benini, S. (2018). Hair detection, segmentation, and hairstyle classification in the wild. Image and Vision Computing, 71, 25-37. (link)
