How to select points of object in dense point cloud?

[morioka7]

Hello,

I try to following process by command-line interface.

• Generate dense point clouds by using series of image
• Select points of objects (such as a vehicle) which appears in a image from dense point cloud.
• Export selected points (such as a vehicle) to file.

Is it possible by using COLMAP? I think "Consistency Graphs" is one of possibility. Does "Consistency Graphs" have mapping information from each dense point to pixel in the image,

Thank you in advance for your suggestion.

• Morioka

[tsattler]

The consistency graph will give you for each 3D point the list of images from which it was generated (see https://colmap.github.io/format.html#consistency-graphs for details). You can then project the 3D points into the images (using the intrinsic and extrinsic parameters in the sparse/ folder) to get the corresponding image pixels. You can then use these pixels to determine whether the 3D point corresponds to a structure, e.g., vehicle, that you want to export. However, you will need to implement this yourself as there is no native functionality for this in COLMAP.

[morioka7]

Dear tsattler

Thank you for your response. I have additional questions.

You said “project the 3D points into the images”. Does these 3D points mean dense point cloud or parse point cloud?

Thank you.

-Morioka

[tsattler]

Dear morioka7,

from what I understood, it seems you want to find out which of the dense points belong to objects. In this case, it seems you would want to project the dense points into the images.

[morioka7]

Dear tsattler

Yes, I would line to project the dense points into the images. Is it possible by using "Consistency Graphs"?

-Morioka

[tsattler]

As described above, you can use the consistency graphs to determine which dense point was created from which image.

[morioka7]

Thank you for you help. I will try consistency graphs.

[fclairec]

Hi @Morioka7, Hi @tsattler I would like attempt something similar. Could you give me some help with opening the consistency-graph files? I would like to stick to python for doing this. The description sais:

the graph is stored as a mixed text and binary file, while the text part is equivalent to the depth and normal maps and the binary part is a continuous list of int32 values in the format .... Here, (row, col) defines the location of the pixel in the image followed by a list of N image indices.

I am using width, height, channels = np.genfromtxt(fid, delimiter="&", max_rows=2, usecols=(0, 1, 2), dtype=int) but can't figgure out how to get the binary part.

Also I would lik to ask you if the consistency graph is really able to give me the following mapping: 3D point to image indexes used for reconstruction. The description just above does not include this, rather (row,col) seem to be locations on the images themselves.

I would be really greatful for some pointers here. Thanks a lot :)