Page 2 of 2
Posted: Mon Jul 16, 2007 4:01 pm
by Nik
Maybe we can use this thread for further discussions on computer vision in general, as you can read in the Linux section I'm experimenting with webcams myself at the moment.
Here are some gerneral thoughts I have to the subject:
I think that for understanding machine vision we must understand natural vision and I have found some conclusions mentioned below alongside an example from natural vision.
1. Colors are not important for object recognition. Example: bw photography/movies and animals with bw vision
2. Resolution is not very important for object recognition. Example: insect eye
3. Stereo vision is not essential: Example: One-eyed pirate
4. Contrast is important for object reognition. Example: looking through milky glass is hard
Posted: Mon Jul 16, 2007 4:14 pm
by ricardo
Great!! I will test it and study the code
Again, many thanks for your time and effort
Posted: Mon Jul 16, 2007 5:09 pm
by PureLust
ricardo wrote:Again, many thanks for your time and effort
You are welcome.
Nik wrote:1. Colors are not important for object recognition. Example: bw photography/movies and animals with bw vision
2. Resolution is not very important for object recognition. Example: insect eye
3. Stereo vision is not essential: Example: One-eyed pirate
4. Contrast is important for object reognition. Example: looking through milky glass is hard
at 1. - right - but the aditional information makes it easier.
at 2. - insects hardly can identify objects - but they can identify movements very well.
at 3. - right - othervise TV won't work.
at 4. - Contrast is not this important for a computer as it is for a human eye. I'ts no problem for a machine to differ exactly between 256 (or even more) greyscales but for a human eye it's nearly impossible to differ between each of them.
If you ever get flashed by the Police because you've driven to fast and looked at the Snapshot they send you, you know what Computers can do with contrast.
Looking through a milk glass is in someway compareable to the view an insect has - it reduces the resolution tremendously.
It let you recognice movements but hardly let you identify Objects.
Nik wrote:... as you can read in the Linux section I'm experimenting with webcams myself at the moment.
Yes, I had a look at this before.
Good luck with that.
Posted: Mon Jul 16, 2007 5:31 pm
by PureLust
@ricardo ... just a Tip to one of your other projects:
I've seen your Thread "
Moving through 3D game using webcam, small example" and the pictures where you control something with your hands.
In an earlier post ricardo wrote:... but may have more than one object with same color ...
If you want to identify multiple objects whithin the same colorrange and the area of the objects are limited as they are in the picture above, you can still use the routine to do that.
Just cut the WebCam Image into a left and a right part and analyse each of them by the routine. You will get the position of an object in the left and in the right half of the WebCam Image.
E.g. set the colorrange to the color of your hands, and you will able to control by your hands (as long as there is nothing else in the Image which will fit into this range).
Posted: Mon Jul 16, 2007 6:22 pm
by ricardo
PureLust wrote:
If you want to identify multiple objects whithin the same colorrange and the area of the objects are limited as they are in the picture above, you can still use the routine to do that.
Okay, thanks as usuall.
But in fact there are differenr projects. Well, the webcam games and experiments was just that, experiments of a sunday, not real projects, at least until now. (Im getting interested on all this webcam stuff however)
The detection is a project of mine and i have to detect in all the screen, so i guess i need to use different colors.
Posted: Mon Jul 16, 2007 9:17 pm
by Nik
Couldn't it be possible that if you generate a histogram for the color you search for looks like that:
Code: Select all
/\
/\__ / \
/ \ / \
_____/ \________/ \______ x-axes
where every hill repesents a high density of pixels within the range to find more than one object when combining histograms of x and y axes?
Posted: Mon Jul 16, 2007 9:40 pm
by PureLust
Nik wrote:Couldn't it be possible that if you generate a histogram for the color you search for looks like that:
Code: Select all
/\
/\__ / \
/ \ / \
_____/ \________/ \______ x-axes
where every hill repesents a high density of pixels within the range to find more than one object when combining histograms of x and y axes?
Yes, you could identify multiple Object in the x-histogram and although in the y-histogram, but because you're missing the 3rd-dimension you cannot recontruct which hill depends to which Object.
E.g., you have both x/y-histograms looking like the above, you cannot recontruct if there are 2, 3 or 4 Objects or even where exactly the are.
Code: Select all
------- ------- ------- -------
| O | | O | | O | | O O |
| O | | O | | O O | | O O |
------- ------- ------- -------
For all the above examples the created histogramms will look like yours.
Posted: Wed Jul 18, 2007 5:57 pm
by PureLust
I tried my code under Linux and find out, that some of the used structures are not predefined under Linux.
So I've extendet the code in that way (see source and links above).
The routine seems to be much faster under Linux than it is under Windows which could be foundet to much faster Point-/Plot-Functions under Linux.
If the Speed under Windows is a problem for somebody, it could be speeded up by about 4-5 times under windows as well using some API-Calls.