DeDoDe doesn't produce absolute confidence?

Open nnop opened this issue 1 year ago • 2 comments

Hi, thanks for the great work.

I found that DeDoDe calculates the Softmax confidence over the entire image and the max resulting confidence is very low, e.g. 0.00095387973 in my test. https://github.com/Parskatt/DeDoDe/blob/0928debd09de6e814c90647a639a7219437bf49a/DeDoDe/detectors/dedode_detector.py#L47

So how can I set an absolute confidence threshold (for example: 0.1) to filter out non salient area?

Apr 24 '24 07:04 nnop

I don't recommend doing that. The filtering is done by topk. If you want to set a threshold, perhaps you can set it based on the max over the image, e.g.., t= 0.01 * maxprob

Apr 24 '24 07:04 Parskatt

Thanks for your suggestion. I think that's practial.

Another question about the params. Why set use_nms=False in detect(). Wouldn't nms help?

Apr 24 '24 08:04 nnop