hello

[spacex22]

hi is the system working ? and what camera does it use ? can it capture point cloud from video ? may you tell the build setup ?

[morsingher]

Hi, I will answer to each question separately:

1. Yes, it is working. Just stick to the main branch, as the others are either experimental or customly modified for specific projects.

2. Currently, only the pinhole model is supported and the unprojection to 3D is done with the usual equation: 3D_point = depth * inv_K * pixel. More in general, this repo is meant to fuse depth maps from MVS, therefore the input is assumed to be in the same format as for MVSNet (https://github.com/YoYo000/MVSNet#file-formats), while depths and normals are loaded with the COLMAP format (https://colmap.github.io/format.html#depth-and-normal-maps).

3. The code can build point clouds from any number and any kinds of depth maps. If you want to generate a point cloud from a video, you need to compute depth and normal maps for each frame with MVS (i.e. COLMAP or any other system), then you can run this code to obtain the point cloud.

4. The build setup is very simple. If you have OpenCV and PCL installed, all you need to do is the usual CMake steps: mkdir build && cd build cmake .. make

As a reference, the following result is obtained with this code on a long video from the KITTI dataset:

Hope this helps!