I’ve just looked into the fitness function, which defines a weight to prioritise the different metrics when choosing the best checkpoint. The default for [P, R, map@50, map@50-95] is [0, 0, 0.1, 0.9].
As I understand it, this makes sense for large to medium sized objects, as we prioritise tighter bounding boxes (which imply better map at varying thresholds) and hence the default weight.
I am working on very tiny objects (~12x12 in a FHD image), and would like to change the weight to only prioritise map@50 or less. This is because I suspect that the model has multiple cases where the IoU fluctuates massively. Obviously, I can directly edit the source code in my package files for now, but wanted to know if there was a yaml argument / commandline argument that could do this for me.
I checked this PR: Add support for defining custom fitness metric by arjunchandra2 · Pull Request #18269 · ultralytics/ultralytics · GitHub, but it seems to have been closed without merging it.
TIA.