Face/Headsegmentation labeling tool
We are releasing a free public version of our facial/headsegmentation tool to allow to generate fully pixel-level segmented facial regions. You can use this tool to help increase our face/head segmentation dataset to gain access to it.
The tool comes in two flavours:
- A version that requires CUDA/CUDNN installed. This version uses a trained segmentation neural network to generate a first segmentation this tends to be a good starting point and often only requires a bit of cleaning up and fine-tunning
- A version without CUDA/CUDNN installed. In this version the hard works needs to be done mostly from scratch
Both version come with a variety of tools that are on the left hand side of the user interface. (See screenshot below). Both version also allow you to generate some common facial landmark polygons automatically this can helps to speed up the process significantly.
To start the process you can either press Load Dir(ectory) or Load Image button. Load Directory means it will load all the images it can find in the directory and you can go through them one by one by pressing the Next (just skip to the next one) or Save/Next (save the current label image) and move to the next one. When using Save/Next the image and the lable images are automatically stored in a PROCESSED directory and moved away so next time you use Load Dir button it will not be loaded again.
To mark areas of the image with a label you basically create label polygons you choose the label you want to create and then press create Label Poly. Each time you click with the mouse in the image area you create a new vertex of the label polygon you can finish the label polygon by either using right mouse click or adding another vertex on the same place of the first vertex.
After a while you will get a picture like this
As you can see from the screenshot each created label polygon is listed in the right area of the screen. So you can always come back to a previous label polygon delete it if you do not like it or smooth it.
With the move up/down/left/right the current label image can be positioned in relation to the source image. The segmentation network does not always create a label image that has the same dimension than the source image is is due to the encoder/decoder structure of the network. So that is why it can be slightly shifted and not line up correctly.
With flatten/bake polys you bake the current label polygons into the label images be careful doing this as these label polygons will then disappear and can no longer be altered
Mouse wheel can be used to zoom in / out of the image at the current mouse cursor position very handy when tracing a particular edge
- Linux (Coming work still in progress)
- MacOS X (Coming work still in progress)