The current GVI calculation doesn't seem to a good solution (at least not yet). See #29 (comment)

Something seems off (either with the methodology or with the code) based on a spot check of the 3 highest and 3 lowest GVI scores.

Additionally, it's likely that there are better ways to analyze than the green view index calculation from the original Treepedia. Can we download and run a transformer model to do class/semantic segmentation and calculation the % of pixels classified as a "vegetation" type category? (Noting that we want to run locally and not depend on an API/service for the analysis.)

For example https://huggingface.co/iammartian0/RoadSense_High_Definition_Street_Segmentation can segment out Nature-vegetation as a category.

objective

easiest option at this time would seem to be to borrow and build on what the creators of https://github.com/Spatial-Data-Science-and-GEO-AI-Lab/StreetView-NatureVisibility have done and...

• crop the bottom 20% band as our images will likely have a car roof or helmet like this:
  
• Apply semantic segmentation to assign labels to different regions or objects in the image.
• I'm not sure why they say "Divide the image into four equal-width sections." after the segmentation. my assumption was that the warping visible in the image above capturing the full scene might impact the accuracy of the segmentation? do we want to divide the image into 4 images with 90 degree field of vision instead of 1 image with 360 field of vision before the segmentation step?
• calculate the % of the 360 image that is vegetation/trees/etc
• attach the score to each point
• save out the geo-file following the conventions of the existing steps
• if possible, consider having an argument to save out the X number of highest and lowest GVI scoring images to let the user spot check the results? in a format like they do:
  
• see the sample segmented image above. there's a more pastel green (grassy looking area) and a more neon green (appears to be mostly trees). confirm the different possible labels related to vegetation (there are at least 2) and check back to discuss what we want to include.

other notes

• this would be an additional file alongside assign_gvi_to_points.py - it would be an alternate script to run for that step. It would not replace that file. The script should read in a geospatial data file from step 2, check each point for data about an associated image, read the image from disk and analyze it, write the details of the analysis to a new column and save out a new geospatial data file.
• the readme instructions should build on our existing readme.
• this process should be a simple swap for our treepedia based option in 3. Assign a Green View score to each image/feature - when @dragonejt in fix: add LocalImages Image source to assign local images to points using EXIF data #42 is tackling having 2 options for step 2, he has gone the route of including an argument for the method. that is, when running python -m src.download_images the user can include a MAPILLARY or LOCAL argument. similarly, for step 3 calculating green view, we may want to let users include an argument defining their preferred analysis method, something like TREEPEDIA or MASK2FORMER