weecology / EvergladesSpeciesModel

A deepforest model for wading bird species prediction.
MIT License
1 stars 1 forks source link

expand the bounding box slightly. #18

Closed bw4sz closed 1 year ago

bw4sz commented 2 years ago

If we can add a small buffer before we crop the species class boxes, we can recover any cut-off birds.

ethanwhite commented 2 years ago

@skmorgane & I visualized ~20 images and a bunch of different additional buffers and we think that moving from the current 50x50 to 60x60 represents a reasonable compromise between making the boxes big enough to typically include all of the large birds while introducing too much additional overlap between nearby WHIB's.

ethanwhite commented 2 years ago

That said, there are still definitely cases where this isn't big enough to capture entire large birds, e.g., the GREG in this example Screenshot from 2021-11-30 11-41-29

ethanwhite commented 2 years ago

Increase bounding box size from 50x50 to 60x60 or 70x70 doesn't appear to have any significant impact. Decreasing it to 20x20 does negatively impact performance (confirming that we're actually doing something with the changes).

There are tradeoffs between smaller and larger species, so one option might be different box buffers depending on the species (or ultimately, bounding box labels instead of point labels).

ethanwhite commented 2 years ago

Doing one more run at 100x100 and then probably calling this good.

ethanwhite commented 1 year ago

No improvement in any of these experiments.