Limit on the number of detections??

[phj1012]

Is there a limit to the number of detection results?

I wonder if there is a limit to the number of detection results on TextFuseNet.

there are a lot of things that I can't catch even though I increased DETECTIONS_PER_IMAGE and other option In the config/default.py file,

I used the model that is reset 101 backbone as it is.

Is there an output limitation like a network using RPN in the past?

If there are not many objects in the image, I'm very good at catching them, but if there are many objects, I can't catch them.

So I wonder if there is a limit to the number of detection results in the network.

[kczzq]

I had the same problem.
Changing parameters does not solve the fundamental problem.
There might be limitations in output due to the network, so try using a different network.
maximum output calc rpn feature shape x anchor size. That's my guess.

[lehoangHUST]

hello @kczzq, @phj1012 Do you have solve problem??

[ddemilla]

I struggled with this one too. If you have a large number of objects (>50), the confidence scores are likely very low. When running the demo, you can see these detections with --confidence-threshold=0.0 if you also set DETECTIONS_PER_IMAGE to a high number. However, this will lead to a large number of false positives and undesirable detections as well.

I haven't figured out how to get good confidence scores with a large number of detected objects yet (40-100) but you will at least be able to see the additional detections using the above approach.