Multi View Stereo - Benjamin Humphrey monument

Set of 18 frames (1080x1920 pixels) for which we want to build a dense 3d reconstructuion (point cloud).

We are gonna use Structure from Motion 10 (SfM10) to extract the camera poses (Structure from Motion) and Multi View Stereo 10 (MVS10) to reconstruct the dense 3D scene (Multi View Stereo).

Animated gif showing the dense 3d reconstruction built by MVS10 using a downsampling factor of 4 and a minimum number of image points of 3. There are 1,650,998 3D points in the dense reconstruction.

Input "mvs10_input.txt" used:

duh.nvm
100
0.5
32
0.9
30.0
10.0
4
0
4
1
0.5
3
2.0
1

The downsampling factor for the depth map computation was set to 4, meaning that the rectified images for each camera pair were downsampled by a factor of 4 in each dimension prior to computing the depth/disparity maps (just to be faster). Let's see what happens when the downsampling factor is set to 2 instead (it should take longer but the depth/disparity maps should be more accurate).

Animated gif showing the dense 3d reconstruction built by MVS10 using a downsampling factor of 2 and a minimum number of image points of 3. There are 1,610,097 3D points in the dense reconstruction.

Input "mvs10_input.txt" used:

duh.nvm
200
0.5
32
0.9
30.0
10.0
4
0
2
1
0.5
3
2.0
1

Not a whole lot of difference for a much longer cpu time. Note that for some unknown reason the minimum number of matches was also changed from 100 to 200. No big deal.

For both runs, the minimum number of image point matches (for each 3D point) was set to 3, meaning that you needed each 3D point to be seen by at least 3 cameras. Let's see what happens if we set the minimum number of image point matches to be equal to 4. We should obtain fewer 3D points but a more accurate 3D scene reconstruction.

Animated gif showing the dense 3d reconstruction built by MVS10 using a downsampling factor of 2 and a minimum number of image points of 4. There are 852,847 3D points in the dense reconstruction.

duh.nvm
200
0.5
32
0.9
30.0
10.0
4
0
2
1
0.5
4
2.0
1

As advertised, the dense reconstruction has fewer points but it is more accurate.