Add downsampled version of leaves method. #2
Description
Related to NG's Chunk based supervoxel loading.
Neuroglancer now only requests chunks within the current field of view. But if the field of view is huge, i.e. showing a zoomed out dataset, the graph server gets flooded with requests. And worse: The graph server has to load all the respective Lvl1 chunks from disk.
Possible solution: Store "important supervoxels" on each layer of the chunked graph. Important supervoxels could be the ones that are visible for a given LoD of the watershed segmentation ... Even that might be too much once every important supervoxel occupies only 1 screen pixel.