Why is tlt-infer’s label confidence threshold high above 1?

Hello! I’m trying to ask you something.
I don’t know why is this…?

Why is confidence threshold value above 1~.

Why detected Big bbox with nothing object?
It’s confidence threshold value is 1~49…?
very high confidence

Why is there a value greater than 1?
=> 8.908, 8.738 , 8.135 , … ?

This is tlt-infer [images_annotated] (KITTI format).

person 0.00 0 0.00 541.394 277.975 568.149 372.288 0.00 0.00 0.00 0.00 0.00 0.00 0.00 8.908 <-
person 0.00 0 0.00 603.314 118.128 631.804 224.751 0.00 0.00 0.00 0.00 0.00 0.00 0.00 8.738 <-
person 0.00 0 0.00 275.850 132.499 301.142 203.275 0.00 0.00 0.00 0.00 0.00 0.00 0.00 8.135 <-
person 0.00 0 0.00 72.996 146.577 100.588 229.443 0.00 0.00 0.00 0.00 0.00 0.00 0.00 7.638 <-
person 0.00 0 0.00 549.187 92.917 569.934 153.916 0.00 0.00 0.00 0.00 0.00 0.00 0.00 7.625 <-
person 0.00 0 0.00 490.815 65.638 511.604 120.052 0.00 0.00 0.00 0.00 0.00 0.00 0.00 6.988 <-
person 0.00 0 0.00 275.135 196.468 305.882 289.367 0.00 0.00 0.00 0.00 0.00 0.00 0.00 5.023 <-
person 0.00 0 0.00 247.067 126.604 273.698 197.850 0.00 0.00 0.00 0.00 0.00 0.00 0.00 4.982 <-
person 0.00 0 0.00 583.974 267.279 616.247 371.981 0.00 0.00 0.00 0.00 0.00 0.00 0.00 4.646 <-
person 0.00 0 0.00 462.027 51.587 484.034 103.122 0.00 0.00 0.00 0.00 0.00 0.00 0.00 3.966 <-
person 0.00 0 0.00 109.015 129.741 132.023 194.173 0.00 0.00 0.00 0.00 0.00 0.00 0.00 2.856 <-
person 0.00 0 0.00 100.377 0.000 187.369 210.873 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.876 <-

Please help me

Which network did you run?

I used detectnet_v2_resnet18.

How about the tlt-infer result? Is it correct?

Yes. I think the reasoning results are correct.
But there is a false detection.
And the confidence value of this false detection is higher than expected.

The resolution of this picture is 640x480.

This is the label data for this image.

person 0.00 0 0.00 305.651 71.320 320.978 125.177 0.00 0.00 0.00 0.00 0.00 0.00 0.00 6.624
person 0.00 0 0.00 276.846 134.464 298.805 197.074 0.00 0.00 0.00 0.00 0.00 0.00 0.00 6.587
person 0.00 0 0.00 274.163 198.808 305.418 286.720 0.00 0.00 0.00 0.00 0.00 0.00 0.00 5.929
person 0.00 0 0.00 306.533 116.291 532.249 480.000 0.00 0.00 0.00 0.00 0.00 0.00 0.00 5.623
person 0.00 0 0.00 67.477 149.437 265.098 480.000 0.00 0.00 0.00 0.00 0.00 0.00 0.00 4.968
person 0.00 0 0.00 549.517 93.632 570.747 150.209 0.00 0.00 0.00 0.00 0.00 0.00 0.00 4.933
person 0.00 0 0.00 605.537 115.412 633.826 203.251 0.00 0.00 0.00 0.00 0.00 0.00 0.00 4.923
person 0.00 0 0.00 488.322 62.899 512.276 120.353 0.00 0.00 0.00 0.00 0.00 0.00 0.00 4.404
person 0.00 0 0.00 463.915 45.860 479.565 94.439 0.00 0.00 0.00 0.00 0.00 0.00 0.00 3.752
person 0.00 0 0.00 535.929 276.231 575.838 372.922 0.00 0.00 0.00 0.00 0.00 0.00 0.00 3.739
person 0.00 0 0.00 77.547 160.313 97.522 226.982 0.00 0.00 0.00 0.00 0.00 0.00 0.00 3.605
person 0.00 0 0.00 250.061 129.171 273.527 190.192 0.00 0.00 0.00 0.00 0.00 0.00 0.00 3.571
person 0.00 0 0.00 311.867 70.733 324.817 114.261 0.00 0.00 0.00 0.00 0.00 0.00 0.00 2.988
person 0.00 0 0.00 337.746 63.603 348.843 102.257 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.026

I want to know two things.

  1. Why is the confidence threshold value above 1 after tlt-infer?
  2. Why was that big box that was meaningless in the image detected? This confidence values are 5.623 , 4.968.
    These values ​​are higher than expected…

Thanks for the details. I am running default jupyter notebook against KITTI dataset to check.

For 1), I can see similar result. I am syncing with internal team.
For 2), That is related to your tlt model. The higher mAP, the less FPs.From the result of training KITTI dataset on my side, I find several pngs which have FPs, but the score is often relatively lower than TPs.

Thank you Morganh. I will wait for the answer.

@h3a1n7
For 1), tlt team plans to handle it for next release.
This issue should only occur at detectnet_v2. Other networks have not this problem.