scienxlab / unmap Goto Github PK

Poco, Mayhua & Heer 2018 use DBSCAN. Seems like it could be slow. Not sure can parameterize it for the general case either (easier for the coloubar only)

Now sometimes works without knowledge of color map.

Add something on similar tools/work.

Maybe no need to go full MyPy, but some hints would add to the docs.

Instead of Euclidean distance in RGB, try other metrics, esp CIEDE2000 as used in Poco, Mayhua & Heer (2018).

They also use pixel patches around target pixels as a way to smooth the colours a bit. Could find targets based on clustering to do this.

Instead of trying to sparsify the graph of colours, possibly this dimensionality reduction step will work.

This paper uses Laplacian eigenmaps: https://ieeexplore.ieee.org/document/9395231 (GitHub: https://github.com/yuanlinping/deep_colormap_extraction)

They are following the recipe in here: https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.12567

Some Python resources on the topic:

• https://juanitorduz.github.io/laplacian_eigenmaps_dim_red/
• https://github.com/JAVI897/Laplacian-Eigenmaps
• https://pythonawesome.com/fast-laplacian-eigenmaps-lightweight-multicore-le-for-non-linear-dimensional-reduction-with-minimal-memory-usage/

Allow expression of log scale for the colourbar, along with vrange

Inline docs have some TODO etc, eg https://github.com/scienxlab/unmap/blob/main/unmap/unmap.py#L252

Could also add defaults and either type hints or explicitly mention in args/returns.

Should be able to pass an image file (or URI for one) on the command line. The result should be a CSV file of the data.

E.g.

unmap --crop 115,72,690,647 --cmap jet --output data.csv image.png