Validate 3D metrics - #261
Conversation
Sorry, something went wrong.
| shellingham = "^1.4.0" | ||
| scikit-learn = ">=0.24.0" | ||
| Shapely = ">=1.7.1" | ||
| Shapely = "^1.8.1" |
There was a problem hiding this comment.
@gatli didn't we run into problems with Shapely at some point?
Sorry, something went wrong.
There was a problem hiding this comment.
Yup, we can't use it. It doesn't have wide enough binary support.
Sorry, something went wrong.
There was a problem hiding this comment.
So what should we do about this? @Anirudh-Scale how much do we rely in the shapely dependency? Any chance we can get around this? If not, @gatli how can we still make this work?
Sorry, something went wrong.
There was a problem hiding this comment.
@pfmark I can implement a function to calculate the area intersection without Shapely, it's just going to be significantly slower. If that's a trade-off we're ok with, I can do that.
Sorry, something went wrong.
There was a problem hiding this comment.
There's already a function for polygon area intersection in the directory under geometry.py. I'd suggest using that since it looks like you're using shapely only for polygon intersection area.
Edit: it looks like the native polygon was used originally before changing to shapely? The native polygon intersection function is quite slow since it's written in Python, but there's a speed up for convex polygons
Sorry, something went wrong.
pfmark
left a comment
There was a problem hiding this comment.
few smaller comments. In addition, do we have a way of easily testing the correctness of the metrics in a test unit test? I think that would be great!
Sorry, something went wrong.
|
|
||
| def __init__( | ||
| self, | ||
| enforce_label_match: bool = False, |
There was a problem hiding this comment.
This argument (and maybe confidence_threshold?) should be required in the constructor, and if we want to set defaults then set them in the child classes.
Sorry, something went wrong.
Reviewers
At least 1 approving review is required to merge this pull request.
Added 3D IOU, BEV 2D IOU, and recall, precision metrics. Tested end-to-end with celery in models repo (see associated PR there).