Entity Localization Bug: Symbol. Bounding box includes parts of figure in paper 1901.10159v1

[andrewhead]

Description: The bounding box of some symbols stretches to include pixels from figures. Here is a case of this happening for several symbols all on the same page.

URL (optional): If you run a local version of the user interface connected to the dev_symbol_failure schema, you can observe this issue at http://localhost:3001/?file=https://arxiv.org/pdf/1901.10159v1.pdf&preset=demo

How to fix (optional): There are at least two potential fixes to this issue:

1. Outlier detection: exclude bounding boxes that are this tall
2. Decolorize figures: Replace figure assets with equivalent-size images that have no hue (i.e., are just black or white). That way, when our entity localization code runs, it will not detect the colors within the figure.
3. (Unknown) There's something strange going on that this bug occurs because no colors should be detected in the figure at all, because the figure should stay the same before and after the colorization. So the first task in debugging this issue should be in determining why the figure appears in the image diffs between the uncolorized and colorized versions of the paper, and seeing if there is a way to prevent that figure from appearing in the diff.

[andrewhead]

Note that sentence bounding boxes also expand to include parts of the figure as well.