Calc the Seconds moments of area of huge number of countours

edit

link text I've cut a litle piece. I want to calculate the volumes and the second moments of intertia of each triangle. I've put some numbers just to have an example. So, the number 1 has 0% of Volume, it's empty, the seconds moments are null. The number 2 has 100% of volume, second moments would have big values. Number 3 would have 60% of volume more or less, and the second moments would be intermediate. Number 4 it's a big hole, I don't want to calc anything.

I think it's clearer what I'd like to calculate.

Ok. So you need to do some pre-processing on your triangle mask. Threshold it, so you have lines of 255 separating your black triangles. Then do a bitwise_not to invert that. Now you have a separate areafor each. Run connectedComponents to get a different label for each area.

For your other image, convert it to a binary image.

Now, one by one, use inRange to extract a single triangle from the label image as a mask. Then do a bitwise_and with your floorplan layer. This leaves the only the area that is above zero within that triangle. Now call moments with the binary flag on, and that's the results for this triangle. To calculate percentages, you'll need to also use moments on the triangle mask after you've extracted just one triangle, to compare to the total area.

@Tetragramm It's not easy. Should I keep all the mesh contours in white? Or some in white and others in black? Could I use other color to all mesh, and then calculate moments?

another alternative approach. If I have the mesh and the "layer" where the mesh is. Is it easier to isolate? Mesh image: https://dl.dropboxusercontent.com/u/7... Layer image: https://dl.dropboxusercontent.com/u/7... code: http://pastebin.com/Q3sN5akv

@Tetragramm is it something like this that is need to calc the moments per triangle? https://dl.dropboxusercontent.com/u/7...

Not quite. Take your Mesh image there, and use Connected Components to get a label for each triangle.

Then extract just one triangle from the image.

You should have a single white triangle, and the rest black. Make a copy of your layer image and use bitwise_and with the single triangle to erase everything but what is inside that triangle.

Now you have a small amount of layer left. Calculate the moments, and that gives you one triangle.

Repeat for every connected component.

@Tetragramm the connectComponents function is not available in python :\ Any alternative, or I should just implement in C++ ? It won't be easy to me, I'm used to use opencv python

It should be available. It's used in THIS python tutorial.

@Tetragramm you've said " Make a copy of your layer image and use bitwise_and with the single triangle to erase everything but what is inside that triangle." How can I select just one triangle for each time. I think that I'm understanding the process. Having n-masks, one for each triangle. Then for each mesh-triangle I can calc moments, centroids, etc . However. I don't get the "process" to reach it.

Well, in the label image that returns from connected components, each triangle has a different label. You can use several methods, the simplest of which is compare.

There's an easy way to call it: Mat singleTriangle = (everyTriangle == labelNumber);

when I appy the ret, markers = cv2.connectedComponents(imMeshly_bw), with imMeshLayer = cv2.imread('testLayer6_.png', -1) imGrayMshLayer = cv2.cvtColor(imMeshLayer, cv2.COLOR_BGR2GRAY) (thresh, imMeshly_bw) = cv2.threshold(imGrayMshLayer, 128, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU) I got : https://dl.dropboxusercontent.com/u/7...

That doesn't look like THIS image from a couple of comments ago. Whatever you did to get this one, that's what you want to run connected components on. In case you are wondering, there are 3079 connected components, so that should be your maximum label number, with connectedComponents returning 3080.

Ok, I think that I'm understanding what you are saying. This is my current code: http://pastebin.com/HfuGUD2N I have 3080 connectComponents. But I don't know how to deal with the information

It seems odd to me. I have more than 3080 components. There are far more triangles.

Now use whatever method you prefer to get an image with just one of the labels (see 4 comments ago). Then use that as a mask to copy your other image, so the only part remaining is inside that triangle. Then do moments.

@Tetragramm something is wrong. I look into the labels matrix. It's size is the same of the original image. However, they are 3079 different connectedComponents . If select according to the label value, I select a lot of triangles not only one.

@Tetragramm when I use _, contours, hier_ = cv2.findContours(imMesh_bw1,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE) cv2.drawContours(colorMask,np.array(contours),-1,(0,0,255),6) it returns 125091 contours (probably the total number of triangles), and if I draw it, it returns this image : https://dl.dropboxusercontent.com/u/7... . Could this be an alternative? i'm not handling properly connectedComponents, I think.

Ah, I see the problem. I thought the mesh layer was just black and white, but it's not. So it's connecting nearby triangles. Instead of using THRESH_OTSU, set a manual threshold of 250. Then connectedComponents returns 13052.

Contours might not work because some of the lines may be too close together and join into one.

The highest number that I get of connect components labels is : 1740 and countours is : 172187 with this config : ret, imMesh_bw11 = cv2.threshold(imGrayMsh1, 128, 255, cv2.THRESH_BINARY) and I get this image: https://dl.dropboxusercontent.com/u/7...

I think that finally I get the correct connectComponents value, 13053 . I still don't know how to do what I need to do :)

However there are more triangles than labels. For istance with label 0, it has 7884676 triangles. With the label 1, 19673, etc.

Do you know how many triangles there are in total? Using a threshold of 250 I didn't notice any triangles that shared the same label.

No, I don't know how many triangle do I have. However, I know that are lot of labels = 0, and labels =1 ; so I can assume that are triangles with the same label. I did the tresh = 250 .

@Tetragramm I've tried another approach, without success. :\ http://pastebin.com/kU5ZaSajhttps://dl.dropboxusercontent.com/u/7...

Labels is the image with each pixel replaced with the label it belongs to. Label 0 is background. So look and see, there's only one triangle with label 1, it just has however many pixels.

@Tetragramm I understood what you said, but i'm not able to see the triangle or how to extract it. I'll look into the docs.

Completly stucked . :\

I added a practice image to my answer. It's got 13 triangles (more or less triangles anyway), and it's small enough you can look at every pixel individually if you have to. Use this to figure out how connected components works.

@Tetragramm thx and I'll try. I think my problem is with "masks" and bitwise operations, to get the images that I want. But I'll do the "tests" anyway.

This is my example code : http://pastebin.com/WvPwebS8 How can I get just one triangle ? I'm tryng selec only the labels which are one and then write a image without success

In c++ it would be (labels == 1). In python it's probably better to use the compare function. triangle = cv2.compare(labels, 1, cv2.CMP_EQ) is how I think it goes. (again, I don't use python, so my syntax may not be right).

For other triangles, replace 1 with the other numbers. Remember 0 is background and it goes up to num_labels-1.

I think that I get it : http://pastebin.com/m6FJnQpi I have the handicap with this code. I'll only see the label 14, because it's the last of the set. But I can do for all of the components. What's the next step ?

Good. Now that you have a mask for each triangle, do a bitwise_and with the binary image of your layer image. This leaves you with just the part of the layer image inside the triangle. Then the moments function gives you what you need, as best you've described it.

hi ! my current code : http://pastebin.com/PgucizcR This are my images https://dl.dropboxusercontent.com/u/7...https://dl.dropboxusercontent.com/u/7... But I have a problem. My "triangles" onde the layer are not "good" Because there black and white parts of the triangle, it depends on the geometry of the layer. Should I do other bitwise operation or use this image https://dl.dropboxusercontent.com/u/7...

No, you do the bitwise stuff after the mask is created.

@Tetragramm I didn't get it :) What bitwise stuff ? I know that i need some oeprations, but with so many comments and misunderstandings (of me ) I don't know what they are :) Anyway, I added features to my code, that I think with your help will be helpful. Code: http://pastebin.com/6TgZKXmD output: https://dl.dropboxusercontent.com/u/7...

In your code from HERE there are only two steps left to do. After you calculate the mask for the label add the line maskedLayer = bitwise_and(mask, imLayer1)

Then copy code from HERE where you calculate the moments and replace where you use the contour with maskedLayer. Also pass true for the binaryImage parameter in the moments function.

@Tetragramm Thank you for your support. I think there is a litle thing to solve. This is the updated code, and I think that's what you've recommended http://pastebin.com/NpjDz7Yw Is not possible do the maskedLayer = cv2.bitwise_and(mask, imLayer1) , because they don't have the same dimensions, So I assumed that was maskedLayer = cv2.bitwise_and(mask, imLayerBw) , but it does not work either. I think this is the "litle thing" to solve. it is almost done !

Ok, I did an update http://pastebin.com/kaC8aVQu and the output is https://dl.dropboxusercontent.com/u/7... I think, something is missing, anyway ! But is close

Something like this, but you need to fix the problem with imLayerBw. I don't know why that wouldn't work. It's a single channel image, with the same width and height as the mesh image, right?

imLayerBw is what your are saying :) https://dl.dropboxusercontent.com/u/7... Your code update doesn't have the contour code. Without the code to take the countour how can we get the moments ?