Face/Headsegmentation dataset

Mut1ny is making part of its head/face segmentation dataset available for free. People can use it freely in their own research, private or commercial application if they want. As long as the original dataset is not made public available elsewhere there is not going to be any restriction in its usage. The reason being we discovered there are so few freely/open available segmentation datasets that containing a significant number of test samples for training decent segmentation algorithms. In this particular domain there are known we are aware of.

Our current dataset contains over 23k labeled images (23596 exactly last measured 03/04/2020) and we hope it is going to grow further also with your help.

The commercial/research partner dataset contains over 39k labeled images as it stands now. Check out our Patreon page for more statistics on the content of the partner dataset and also for a comparison of a trained deep neural network between using the normal and the partner-level dataset

*NEW* As of 02/05/2020 the commercial/research partner-level dataset now includes a new label (all of the previous label images have been updated), the new label is specs & sunglasses and its label value is: 0,128,128. This label will stay exclusive for the commercial/research-partners for the next 3 months.

Here a few example images:


 *NEW* As of 02/01/2019 the dataset also includes bounding boxes of the different facial parts

Bounding boxes example 1 Bounding boxes example 1  Bounding boxes example 1 


*NEW* We have now set up a new Patreon page if you want to support this project in any way.

Here a few example images:

You can download a small representative sample dataset containing 400 fully labled examples which you can download for further inspection if you like.

How to obtain?

If you are interested in receiving the whole dataset there are two options: First one is to provide a small data contribution. Second one is to set up a sponsorship via our Patreon. Both options will help us to enlarge the dataset with new data. 

If you decide to go for data contribution option we currently ask people to provide 50 previously unlabled images to be added to the dataset.  We provide both the unlabled images as well as a tool that can help you to generate these 50 segmented labled images 

Don't want to waste your time training, finding the right architecture or setting up a training curriculum or do not have the resources to do so?

Why not checkout our Faceprocessor it offers a ready to plug-in solution and more  

If you are interested in doing so please fill out the form below

Dataset access request

and we're going to provide you with further details.