Gabor Filter - Fine Hair Extraction v2.1
This location is for Registered Users Only.
Perhaps you need to login or register.
14.0, 13.2, 13.1, 13.0 or later
Linux, Mac, Windows
Fine hair keying using Gabor filtering
This gizmo uses the magic of Gabor filtering to provide a way for detecting elongated features (like fine hair strands).
The node outputs the filtered image in the 3 rgb channels. A strong contrast on the source image will give stronger, better results. If you're using it on green screen for hair extraction, you generally will use the green channel only as it is the one with the strongest contrast between the strands and the background.
Try to play with the settings ! The best result I personally obtained were always by playing with the sliders, trying different combinations.
Then, you can use a merge node with operation 'plus' (or 'screen') to add the hair details to your existing alpha (IBK or anything else)
Using a denoiser before the process is strongly recommended !
I've added a small help text for every control knob. Don't hesitate to hover the cursor on those if you want information on what they do.
Please note : the node can be a bit complex to use, and results heavily differ between different inputs. As a result, I prefer mentionning that this gizmo is still a work in progress, that I will continue developping in the future with the hope of understanding the process better and enhancing the results and the ease of use !
This gizmo comes from an idea presented in Chloe Legendre's SIGGRAPH paper from 2017 'Improved Chromakey of Hair Strands via Orientation Filter Convolution'
Enjoy !
----
v2.1 update :
I realized that the 2.0 gizmo was limited in cetain cases
- Re-added the 'hair lighter than BG' check box
- Added some more controls
v2.0 update :
- Overall upgrade for better results
- Changed user interface for more clarity, with less confusing sliders
Comments
I think that the BG Offset is not linked to the CCs. Not sure if it's supposed to control that knob inside.
RSS feed for comments to this post