oversegmentation (was: find groundtruth of an image)

[EDIT: 20110610 09:49 CDT - merge from duplicate - WDR]

[EDIT: 20110608 09:48 CDT - merge from duplicate - WDR]

Hello to everyone in the community,

I'm working on my thesis, in wich I have to generate gaussian spots and extract the groundtruth in an image.I have overlapping spots and i have to take the ground truth of them.Any idea of how to do this?

[Material from duplicate]

Hello to everyone in the community, I'm trying to extract the boundaries of an image with gaussian spots. There are overlapping spots, that with the watershed method they seem to be oversegmented! Any idea of how to avoid oversegmentation. I have found a way with im imhmin but it doesn't work. Here are the images I've extracted. The original http://www.sendspace.com/file/ysyi9y and the boundaries of it http://www.sendspace.com/file/3il5e9

Plz help me I've stuck with it

[Material from duplicate #2]

Hello to everyone. I'm trying to segment some overlapping gaussian spots with watershed algorithm but I've problem of oversegmentation.Is there a method to avoid oversegmentation?

[Walter Roberson]

For those like me who do not recognize the terminology: a description for what groundtruth _is_ seems to be

http://en.wikipedia.org/wiki/Ground_truth

[Image Analyst]

For a synthetic image that you generated, the ground truth is the same as the image, unless you did something to it like added noise or blurred it or something - in other words you have an input image that you degraded somehow and you're looking to restore it. Or maybe you're trying to find the centroid and spread of all the individual Gaussians (like was asked in another thread either here or in the newsgroup recently) from the overlapping Gaussians image. Clearly you need to explain more what you're looking to find.

Answer 1

[Walter Roberson]

I do not understand how your question fits in with what Wikipedia defines Ground Truth to be? Ground truth appears to related to external verification data that an image would have to be brought into line with, but your question seems to imply that groundtruth is something that could be extracted from an image alone?

[Mary Kodogianni]

My supervisor explained me that ground truth is a method to separate overlapping areas and to find the sum of intensities that include these areas. An image of ground truth includes the outlines of my spots but in case i have overapping areas they have to be separated, so if in the original image 2 overlapping spots seem to be one bigger spot, in the ground truth image that i have to extract, i have to take 2 outlines stuck

[Sean de Wolski]

post the images...

[Mary Kodogianni]

I have tryed bwboundaries to find the boundaries of the spots but i don't know how to make them seem to be separated in the ground truth image.

[Mary Kodogianni]

Sorry but how exactly i post the images?I have tryed to submit them

[Sean de Wolski]

http://www.mathworks.com/matlabcentral/answers/7924-where-can-i-upload-images-and-files-for-use-on-matlab-answers

Answer 2

[Mary Kodogianni]

Thanx!Here is the link with the zipped images I have analysed before http://www.sendspace.com/file/y6m16y

[Image Analyst]

No, they're not zipped. I tried to download your .rar file into winzip, and winzip choked. How do I open this thing? Or can you simply upload the original images or a normal .zip file?

Answer 3

[Mary Kodogianni]

Here I've uploaded the original image http://www.sendspace.com/file/15rv3n and the "ground truth image" of the first http://www.sendspace.com/file/4o959z

[Walter Roberson]

That ground-truth image appears to be completely black!

[Mary Kodogianni]

Sorry,by saying ground truth i mean the boundaries-perimeters of the spots,use ImageJ to open the images.

[Walter Roberson]

I don't think I have ImageJ

The file is a 16 bit TIFF. The image editor I have handy converts it to 8 bit, and then finds there is only a single color in it -- implying that any variation was in the bottom 8 bits. If that is the case, perhaps you could convert to uint8 and save that as an image and upload that?

Answer 4

[Mary Kodogianni]

Here are the images as 8bit TIFF. http://www.sendspace.com/file/hbmh5l , http://www.sendspace.com/file/eyrbm4

[Walter Roberson]

Ah there, now I can see something in the ground truth file!

[Image Analyst]

It looks like all you did was to threshold your original image and run bwboundaries on it. Why is that ground truth? How are you defining ground truth? And I don't know what you mean by "don't know how to make them seem to be separated in the ground truth image." What's separated from what? All I could see in your "ground truth image" was individual rings that were already separate from each other and anything else.

[Walter Roberson]

I did see some places where the circular-like structures were grouped together in a blob.

[Mary Kodogianni]

Exactly for these structures I'm talking about!In my "ground truth image" my suprevisor says that these structures have to be separated because the groundtruth is a way to find the overlapping percentage you have in an original image

[Vijayalakshmi Sonnad]

hi, i want matlab code to construct ground truth image of an image for the evaluation of various segmentation techniques

Answer 5

[Abhishek Kaushik]

Ground truth is basically the actual data obtained through physical field surveys on the location.

[Walter Roberson]

... and better yet, verified by an expert.

For example, around here, someone looking at a photograph of a tree might identify the tree as "elm", meaning what they probably incorrectly know as "Chinese Elm" which is really "Siberian Elm" Ulmus pumila, but someone else looking at the label "elm" might think the reference is to the "American Elm" Ulmus americana which is a very different tree. The Siberian elm is a fast-growing disease-ridden weed tree will grow just about anywhere, which we despair of getting rid of; the American elm is a large slow growing long-lived magnificent tree that we actively protect; we have the largest mature urban American Elm tree forest in North America.

Correct detailed labeling of what is there can be very important for correct analysis.

Answer 6

[ruchika lalit]

is ground truth image a binary image or grayscale image or something else

[Walter Roberson]

Something else. Ground truth indicates in a pixel by pixel way which class each pixel of the other image belongs to. It would typically be integer class numbers. Sometimes when the question is just whether a particular pixel is or is not something (such as tumour) then it could even be represented as binary. Ground truth is really an image, but it would be common for it to have been stored in a way that could be displayed as a pseudocolor image. It might well be stored as a label matrix.

[ruchika lalit]

then how can we find ground truth

[Walter Roberson]

Ground truth cannot be computed (well, other than in the sense that a DNA analysis or x-ray diffraction might be considered to be computed.) If it could be computed then everyone would just use that algorithm rather than the not-always-accurate algorithms that are currently used to try to analyze images.

Suppose, for example, that the basic image were of the trees in your back yard. Then ground truth would involve a trained botanist or tree specialist visiting your yard and analyzing what variety of tree each of your trees was -- and if the trees overlapped you would also need to take careful measurements of the way the branches ran so that you could decide the exact boundaries.

I have an arborist come in about once a year to my yard. He typically confuses my P. triloba 'Multiplex' (Double Flowering Plum, pink flowers) with P. glandulosa 'Rosea Plena' or 'Sinensis' (Double Flowering Almond, pink flowers). He is trained, yes, but his "Double Flowering Almond" is not ground truth -- he is typically here before the blooms, and fails to take into account that Double Flowering Almond is shorter than Double Flowering Plum.

My arborist is also perpetually confusing my large fruit tree as being a crabapple, when it is an apple tree, possibly Morden Russet 999. What is the difference between crabapple and apple? Botanically the difference is not the height of tree or arrangement of flowers: the distinction is made according to the size of the fruit, with the cut-off being below 1 inch for crabapple. My tree averages fruit about 1 1/2 inches, up to 2 1/4 inches in better years. But the arborist isn't around when the fruit is ready in early August, so again his trained opinion is not as good as the ground truth that my fruit tree is an apple rather than a crabapple.