All committers have signed the CLA.

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?

1. Ideally, yes - we'd like a bool flag on with_nms.
2. Could you please explain how soft NMS differs from the standard algorithm?
3. Would it be simple to adapt the current algorithm by adding a boolean flag? If so, let's do that.
4. 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?

1. Ideally, yes - we'd like a bool flag on with_nms.
2. Could you please explain how soft NMS differs from the standard algorithm?
3. Would it be simple to adapt the current algorithm by adding a boolean flag? If so, let's do that.
4. 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.

1. 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

2. 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.

4. 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

4. 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

Thanks @YHallouard !

I'll have a look at the start of next week

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.