The geometry based approach is getting relatively close. This would prevent the need to decompress the image slices for a savings of 8x. - [ ] Finish correctly matching contours. This requires improving the inside/outside polygon logic (use winding algorithm?). - [ ] Figure out how to make matching fast (currently O(N^2)). See the https://github.com/seung-lab/crackle/tree/wms_point_cloud_geometric branch.