Thank you Alexey. This parameter is well found by version 2.0.2. I use value = -1
In fact as explained on another post on the forum, I try to make a model of an object photographed in the dark with only 1 light source and in motion, so with a lot of shadows that change. There are many 9000 images.
During my first test without masks, I have a lot of noise and errors. I masked the shadows in the 9000 images and I get a much better dense point cloud because the parts in the shadows are not used anymore.
Now I try to improve again by increasing the number of used neighbors, to test with more. I have many images around but also at different distances. What do you think is the best way to handle this case ?