Open ItsLastDay opened 7 years ago
I've tried simple way to find local maximas in non-overlaping areas (function peak_local_max form skimage). It didn't work as expected.
Next idea is to build kNN graph on top of embedding and find gradients in it.
Clusterisation of tags seems to be a good stuff. Idea: we make clusterisation of tags and show only centers of tags. For now I've tried DBSCAN. I think that hierarchial clustering will do better. With zooming we can show smaller clusters and more tag names.
We can somehow estimate local "crowdness" of area and annotate only one tag from that area. Therefore, not all tags would be annotated at all, but this can be mitigated by #12.