ToPaz ReMask

Started Dec 2, 2009 | Discussions thread
Dan Tong
Contributing MemberPosts: 585Gear list
Like?
The Problem with Extractions
In reply to Toermalijn, Dec 6, 2009

First of all if any of you have tried Remask v1 and found it hard to use, don't worry, Remask2 is a totally different tool as far as ease of use is concerned. Download the free trial version and check it out yourself -but be sure to view the video tutorials so you will get the general idea of the workflow.

Photos in which the subject/object is against a complex background with similar coloring will always be the most difficult to extract (eg. brown object against a brown background). Even our visual system which puts every extract system to shame will have some difficulty in some really tough cases.

Simple uniform backgrounds whose color is very different from the subject/object to be extracted will provide the most optimal conditions for extraction.

However even under near optimal conditions, very fine, very thin hair which has shows low contrast against the background may still be a challenge, especially if the hair curls and overlaps to create "enclosed" areas of background (the "background" or " Do Not Keep" markers cannot be made to flow into these areas automatically).

For this reason, most product demos will choose and use images in which the hair and background can be neatly separated without having to manually mark hundreds of enclosed background areas individually, one at a time, with a very small fine background brush. A favorite image is a dog or other animal where the fur is very extraction friendly : ), rather than a high fashion photo of a woman's head of hair with lots of wispy fine hair with some tight closed curls against a complex multicolored background. Of course when the color of the hair matches the background the task becomes impossible or near impossible.

Keep in mind that when we humans view a photo we have a lot of information available in our brain without which it is impossible to separate the "figure" from the "background". I the case of an image of a head, we recognize that those wispy lines are hair and hair is part of most people's heads, just as we know that a hat is not part of a human head, although in some cases it may be considered part of the "figure" or "subject" rather than part of the background. This type of Information which is not present (not encoded) in the image!

This means that the computer program needs lots of help from the viewer -hence the need for the user to draw an outline of the "subject" or "figure" to distinguish it from the background. Problem arises from the fact that we cannot outline every strand of hair -too much time and trouble. It would be great to possibly mark an example of one or two or even three strands of hair and have the computer learn from that one or two or three examples and do the rest of the grunt work on its own.

A number of plugins do a fairly good job at identifying skin in most photos (e.g Imagenomic's Portraiture for example), but hair is a lot more complicated.

Of course, it would help the computer program to be told some useful characteristics of hair which would work in most cases.

1. Size- hair thickness has a certain value which is maximum and a practical minimum. So we could give the computer a range of values s a proportion of the size of the average head. This would then require that we outline the head area to let the program know that this is the head (and not the eyeball or nose, or neck, or fist etc.)

2. As we trace a single strand of hair from the "head" it gets progressively thinner (it also gets thinner with distance from the focus point).

3. It curves smoothly and never at a sharp angle (no 90 degree corners)

4. It generally does not change color but saturation may change (partially dyed hair may pose serious problems here, as will special lighting and shade in bright sunlight). Highlights will make this invariance less reliable.

5. Hair responds to gravity. Hair strands will not under most conditions lie perfectly horizontal to the upright head for more than a fairly short distance as a proportion of its length. A long strands of hair is likely to orient itself vertically (down) as we trace its path away from the head (there will be exceptions when wind is blowing). We may need to tell the program the direction of gravity for each image.

6. Hair strands will often, but not always, show a parallel trajectory over some part part of their full length.

There are of course special conditions under which the points above may not prove very useful (clumped hair), and there may be other important features of hair which I have overlooked. Whether these observations are practically useful remains to be seen.

Thanks for listening,

Dan

 Dan Tong's gear list:Dan Tong's gear list
Canon PowerShot G12 Sony RX100 Canon PowerShot S110 Sony RX100 II Canon PowerShot S120 +17 more
Reply   Reply with quote   Complain
Post (hide subjects)Posted by
EricNew
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark post MMy threads
Color scheme? Blue / Yellow