0
$\begingroup$

How do I prepare the info of 3D models to use with NN? For example, I have thousands of models with boxes similar to the ones in the image below. I can extract the vertices and their normals that make up the faces of these boxes. Similarly, I would like to prepare the info of the red-shaded surfaces, again I have their vertices and their normals. For future studies, I will have more complex shapes such as cylinders, pyramids,...etc. What would be the best way to represent these complex shapes for NN?

Update: These boxes don't stay in the same position, see the second image I added. I will have different geometric models and different red-shaded areas on the surfaces of these objects. The NN output would be a number for each surface of these boxes/objects. The number represents the surface temperature. The input would be the following:

1- Some climate information such as (air temperature, humidity, ...etc.) 3- The location and size of the buildings that are represented in boxes/or maybe other shapes. 4- The size and the location of the red-shaded areas (red-shaded areas represent the shadow cast by buildings. 5- Material of each surface (concrete, brick,...etc).

enter image description here

enter image description here

$\endgroup$
7
  • 1
    $\begingroup$ Why not as a tensor ? $\endgroup$
    – hanugm
    Sep 14 at 5:07
  • 2
    $\begingroup$ What do you want the NN to look at? A picture is usually treated as pixels; correspondingly a 3D picture could be treated as voxels. $\endgroup$
    – user253751
    Sep 14 at 8:59
  • 1
    $\begingroup$ @hanugm what data goes in the tensor? $\endgroup$
    – user253751
    Sep 14 at 9:00
  • 1
    $\begingroup$ Intensity values in 3D simulation as you told. @user253751 $\endgroup$
    – hanugm
    Sep 14 at 9:44
  • 1
    $\begingroup$ the advantage of voxels is that they are equivalent to pixels, which we know work well. The disadvantage is that they're a lot of data! It would be convenient if the vertex coordinates etc could somehow be passed to a NN but I can't imagine an NN learning to work with that input. $\endgroup$
    – user253751
    Sep 14 at 9:48
2
$\begingroup$

I think, that the answer depends on the application, but a possible choice would be store it as a mesh - a list of vertices $V$ and edges $E$. Instead of edges, one can work with polygons, and define connectivity $F$ - for triplets of vertices $(v_i, v_j, v_k)$.

There is a nice paper on Mesh CNN that can handle various geometric object.

For the special case of boxes maybe there is a more educated approach, but since you would like to handle later more complex shapes, I would suggest to work with this architecture from the start.

$\endgroup$
1
  • $\begingroup$ Very nice paper but the methodology isn't suitable for my workflow. I just updated my post to include more info on what I want the NN to look for. $\endgroup$
    – Julia_arch
    Sep 16 at 5:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.