zinsmatt / spacecarving Goto Github PK

I am beginner in computer graphics and confused about some details of the algorithm.
Here is my overall understanding:

1. create a cube, which should embrace the object we want to reconstruct.
2. using the projection matrix in each view to project the object 3D points into the image plane and get a long array 'fill'. This 'fill' indicate if each pixel in the image plane is in the silhouette.
3. Stack all the 'fill', we can vote for each voxel in the original cube to tell if it belond to the 3D structure we want to reconstruct.

Still, I have some specific question.

1. Why do you subtract the z-axis of pts with 0.62?
2. As far as I know, projection matrix is used to get the image points from the world points. And in your code, the projection matrix is multiplied with the 3D points in the cube cooridnates. So the projection matrix is not like the normal one or the world coordinate is the same as the cube coordinates?
3. Is 'uvs' the projected image points at each view? Why some of the pixels will outside the image boundary after using the projection matrix?

Also, could you suggest some reference to make these cloud points into mesh? Thank you!

Hello sir @zinsmatt how/where can we find the related datasets of other objects as the one used in your project.
along with camera properties and .ppm images

Hello Sir @zinsmatt . I wanted to know how can we visualize the generated .vtr file in 3d software's like blender,meshlab
or how can we convert vtr into obj or stl for easy visualisation

Hi

Thanks for sharing the code . I have some more clarifications about the code.

1. Why did you choose s=120 did you play around with other numbers. I ask because if I want to initialize a voxel grid for my data how do I do that so that it encloses the object. Any tips or educated procedure that I should follow?
2. Why are the x,y,z normalized by max values and why are then the center point subtracted . Can you explain the rationale
3. Why were the points divided by 5 in line 59
4. Again about line 60 how did you decide on the number 0.62. I am asking so that i can do the same with my images.
5. Another question about camera parameters , can you point to the source of the camera parameters. is the camera parameters the projection matrix which is K*[R |t]

Thanks a lot in advance

Is the matrix that is being loaded up a R|T matrix? (Rotational 3x3 matrix with the 1x3 matrix appended to it)

Very intuitive example and many thanks.
Although, from the code, you used only the silhouettes of the images. I tried to generate the shape.txt using 24 occ, like this:

for occ, p in zip(occupancy, pts[:, :3]):
        if occ > 24:
            fout.write(",".join(p.astype(str)) + "," + str(occ) + "\n")

Then, I import the generated shape.txt file into CloudCompare, the result is a bit rough, which is expected, as only silhouettes are used:

Even using the occ > 24, the result is still worse than yours in readme. Could it be something I ignored ?

Many thanks !

Hello.
Is it possible to do this task without the camera properties(projection matrix) which is in the dino_ps.mat file. Or can you suggest me some reference or code which does this task without camera properties as I'm given with the same task of 3d reconstruction with multiple images but no camera properties are given