Points subsampling

[grst]

When I load a SpatialData layer with 1.6M points I get the following warning:

Subsampling points because the number of points exceeds the currently supported 100000.

why is this necessary and would it be possible to load points on-demand as one zooms and pans around as it's done with the images?

[LucaMarconato]

Unfortunately napari doesn't support this and the performance doesn't scale well with many points. The setting change be changed by modifying https://github.com/scverse/napari-spatialdata/blob/main/src/napari_spatialdata/constants/config.py.

@keller-mark started implementing this for Vitessce. We are exploring multiple approaches, one is a Z-curve ordering, described here scverse/spatialdata#974, another one is a reusing the approach used in neuroglancer, described here https://github.com/google/neuroglancer/blob/c1fd0b036198f1f5218c9731bdba4061179887de/src/datasource/precomputed/annotations.md#spatial-index

I implemented it here for 3D points https://github.com/hms-dbmi/tissue-map-tools/blob/91529b2622d370573410aecb51b3b734d50077e3/src/tissue_map_tools/data_model/annotations.py#L926 and it works great in neuroglancer. This can also be read back to a pd.DataFrame https://github.com/hms-dbmi/tissue-map-tools/blob/91529b2622d370573410aecb51b3b734d50077e3/src/tissue_map_tools/data_model/annotations_utils.py#L153 and it would not be much work to write this to Parquet (where chunks and spatial levels are saved as row groups—this information is available as you for instance can see from the last 2 columns of the last output cell from this notebook https://github.com/LucaMarconato/3d-spatial-workshop-2025/blob/main/4.1_tissue_map_tools.ipynb).

So, in principle it would not be a lot of work to do it in napari, but at the moment I'm focusing on improving compatibility with NGFF, so it will be for later (or for whoever may want to take a shot at it).

[keller-mark]

We have an example at https://dev.vitessce.io/?dataset=sdata-xenium_rep1_io that uses the z-ordering described in scverse/spatialdata#974 to achieve tiled loading of millions of points (zoom in to load the points, it may be a bit slow to load. click a gene and use the dropdown menu next to Points in the layer controller to modify the coloring and filtering. shift-click multiple genes in the gene list to view points corresponding to a subset of genes).

• https://github.com/vitessce/vitessce-python/blob/main/docs/notebooks/spatial_data_xenium_morton.ipynb
• https://vitessce.io/docs/data-troubleshooting/#points

As mentioned in scverse/spatialdata#974, an alternative would be to store points using OME-Zarr so they would be inherently tiled / spatially-indexed scverse/spatialdata#789 . Another option for spatial indexing would be to store one parquet table per tile.