Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

How to completely turn off non max suppression? #610

Open
3 tasks
bw4sz opened this issue Feb 20, 2024 · 0 comments
Open
3 tasks

How to completely turn off non max suppression? #610

bw4sz opened this issue Feb 20, 2024 · 0 comments

Comments

@bw4sz
Copy link
Collaborator

bw4sz commented Feb 20, 2024

If we want models to be customizable, we may want to allow users to see all generated boxes. Currently some are suppressed during prediction directly inside retinanet. Then more are suppressed in predict_tile for overlapping images. We want to document a way of turning all of these off. I have a suspicion that the nms settings are baked directly into the retinanet architecture and will be pretty specific. I have often noticed that the config["nms_thresh"] argument works best on the predict_tile methods and not the predict_image, meaning that the nms that is happening inside retinanet isn't being respected. I think the same is true for score_threshold arguments.

  • Document where NMS is being applied.
  • Show an example of turning off NMS for predict_tile
  • Show an example of turning off NMS for underling retinanet model for tree release. This may not relate to other architectures.

Its always hard to guess, but I think this issue will take some careful debugging, since its at the interface of deepforest and torchvision, and not wholly about DeepForest code.

@bw4sz bw4sz changed the title How to completely turn off non-max-supression? How to completely turn off non max suppression? Feb 20, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant