roboflow
feat: Add soft Non-Max suppression
Description
Hi everyone,
I am opening this PR to propose an implementation of Soft NMS based on that paper.
I propose a method with_soft_nms, but I'm not sure that it is in the philosophy of Detection.. Would you have a preference? A boolean in with_nms?
Type of change
Please delete options that are not relevant.
- [ ] Bug fix (non-breaking change which fixes an issue)
- [X] New feature (non-breaking change which adds functionality)
- [ ] This change requires a documentation update
How has this change been tested, please provide a testcase or example of how you tested the change?
I used the same test case than box nms et mask nms.
Any specific deployment considerations
For example, documentation changes, usability, usage/costs, secrets, etc.
Docs
- [X] Docs updated? What were the changes:
Hi @YHallouard 👋
Unfortunately, this will take some time to review. To speed things up, would you have some time to do some of the following?
- Ideally, yes - we'd like a bool flag on
with_nms. - Could you please explain how soft NMS differs from the standard algorithm?
- Would it be simple to adapt the current algorithm by adding a boolean flag? If so, let's do that.
- Could you please provide a Colab that shows the difference in results? Ideally, it should show improvements in the filtered detections.
Meanwhile, I'll accept this as Hacktoberfest-approved. The PR brings the improved algo to our attention, provides the code, a few tests. Hearing what tradeoffs and design considerations were made helps us a lot too! 🤝
Hi @YHallouard 👋
Unfortunately, this will take some time to review. To speed things up, would you have some time to do some of the following?
- Ideally, yes - we'd like a bool flag on
with_nms.- Could you please explain how soft NMS differs from the standard algorithm?
- Would it be simple to adapt the current algorithm by adding a boolean flag? If so, let's do that.
- Could you please provide a Colab that shows the difference in results? Ideally, it should show improvements in the filtered detections.
Meanwhile, I'll accept this as Hacktoberfest-approved. The PR brings the improved algo to our attention, provides the code, a few tests. Hearing what tradeoffs and design considerations were made helps us a lot too! 🤝
Hi @LinasKo, And thank you very much for your review :)
this will take some time to review.
Yes I know haha, this is why I din't started the documentation, I wanted to get your feedback on implementation first.
- Ideally, yes - we'd like a bool flag on with_nms.
Yes, I deeply thought it was the mindset. My concerns was that I need a sigma parameter also, that will be useful only in the case where the bool flag is True. If you have any idea about that, I'll take it :D
- Could you please explain how soft NMS differs from the standard algorithm?
Of course, unlike NMS which removes predictions if they are below a certain IOU threshold, Soft-NMS updates the confidence score of each prediction with a continuous function of the IOU. You can then remove them later with a confidence threshold but the main idea is to preserve overlapping bounding boxes with a good confidence.
- Could you please provide a Colab that shows the difference in results? Ideally, it should show improvements in the filtered detections.
I'll do that as soon as possible, maybe this weekend :)
Thank for your review again
Hi @LinasKo
- Could you please provide a Colab that shows the difference in results? Ideally, it should show improvements in the filtered detections.
Here it is :) Soft-NMS vs NMS
Hi @LinasKo, sorry for the message, did you had the time to look at the notebook ? :)
Hi @YHallouard
Unforunately not. It's on my backlog, but honestly, supervision might steal all the time I have this year.
I'd still like to give it a proper review, so lets see - maybe I can make it happen.