Cluster merging problem

[zhangye1111]

Thank you for your excellent work. Now I have a problem. I ran your algorithm, and when two people are very close, such as 20 cm, the algorithm will identify the two clusters as one cluster. I initially thought it was a cluster merging problem. I set the cluster merging threshold to 0, and even removed the merging function, but the above problem still occurred. Can you tell me how to solve it? Thank you very much！

[Schmluk]

Hi @zhangye1111,

The clustering happens in two steps, first all points are clustered using voxel marching, where points in adjacent voxels are clustered. I think the default voxel size is 20cm, so this is probably why they are all clustered together. In the second step (the cluster merging), clusters that are nearby but don't have connecting voxels are merged to avoid over-segmentation.

In our default use case, clustering of "touching" objects into a single object is expected behavior (we focused more on point-level detection than object-level tracking).

Please feel free to implement your own clustering algorithms (e.g. DBScan could be strong) and add tracking over time (e.g. if there's a Kalman Filter per object intersecting objects would be tracked over time)! Contributions are always welcome!

As a side note: We initially thought about implementing this but the "tracking by detection" performance turned out so good that we ended up skipping these further steps.

Hope this helps!