Holoviews: Significant performance degradation with rasterize(image) * Overlay(objects)?

[zbarry]

Wondering if this is expected given the data I'm working with.

Let's say I have a 2048x2048 image that I have localized ~150 objects in. I then go to plot this image, and because it is large, I rasterize() it with Datashader. Each object is indicated on top of this raster with an hv.Ellipse (which are compiled into an hv.Overlay).

Any sort of zooming / scrolling redraws on the raster are then immensely slowed down if I overlay the rasterized image with the ellipse overlay.

Skeleton example:

hv_ellipses = [
    hv.Ellipse(centroid[1], centroid[0], 4 * centroid[2])
    for centroid in centroids
]

rasterize(hv.Image(image, bounds=(0, 0, image.shape[1], image.shape[0])).options(
    title_format=f'Movie frame with identified objects', 
    colorbar=True
)) * hv.Overlay(hv_ellipses)

I've noted that the slowdown scales with the number of ellipses.

Do you guys think this is expected given the number of ellipses I am drawing? I just realized that in this particular case, I could just draw a scatter plot with transparent faces and circular markers, but I guess I was wondering if more in general (like if I wanted to draw many lines for object trajectories, for example) it is untenable to have many objects overlaid on top of a raster and achieve good performance on raster redraws in response to axes changes.

[philippjfr]

Do you guys think this is expected given the number of ellipses I am drawing?

Definitely not expected, static elements like this should not be redrawn so there should be no performance impact.

[philippjfr]

I can reproduce a pretty large slow down caused by ranges being computed for all 150 ROIs each time, fairly inefficiently. In general it's probably a good idea to replace hv.Overlay(hv_ellipses) with hv.Path(hv_ellipses) but we should be able to achieve a major speedup using Overlays too.

[zbarry]

Thanks for the alternative!

[philippjfr]

Wish I'd recorded the timings here, it does seem to be somewhat faster now at least, I'm getting 0.7 seconds per update.

[zbarry]

@philippjfr - sweet. I'll check this out soon and let you know what I see.

[philippjfr]

Actually getting 420 ms with the latest minor optimization.

[philippjfr]

The main recommendation here is still to combine all the polygons into a single element but the overlay based approach is now also much faster so I'll close this.