pointcloud convert to range image, is it always 64x870? about depth_clustering HOT 3 CLOSED

So, the way it is implemented now was tailored to Frank Moosman's data. These data are all 64x870, which seems to give a good tradeoff between speed and precision. The way I handle multiple points that fall into the same pixel is that only the furthest reading is taken per pixel, but all the points that fall into it get assigned the same label. It can create some minor artifacts, but seemed to work well in practice for me.

You can however, pick any resolution you want if you look into the ProjectionParams class by adding another function like this one with any resolution you want.

Be advised, that working with KITTI data directly might be a little imprecise as the HDL64 has two sets of beams that do not exactly match onto a single depth image.

from depth_clustering.

i think KITTI should be set as:

params.SetSpan(SpanParams(-180_deg, 180_deg, 4000),
SpanParams::Direction::HORIZONTAL);

because Velodyne HDL-64E has 0.09 degree angular resolution.

from depth_clustering.

@niosus

1. I also think HDL64 row step should be 4000 or 2048, but will this increase runtime a lot ?
2. can you explain further about "HDL64 has two sets of beams that do not exactly match onto a single depth image"
3. the AngleDiff::PreComputeAlphaVecs, if the angle is add by step, then it's a constant value, why need a vector to store and obtain it by r,c indices ?

thanks for your work and reply!

from depth_clustering.