my finetuned model is much slower about tensorrt_demos

[jkjung-avt]

This suggested your model probably spends a lot of time doing "NMS". My guess is that your model outputs too many candidate bounding boxes. As a result, the "NMS" op needs to spend a lot time comparing many bounding boxes.

So try the following:

• train your model better, or
• try to use a higher confidence threshold for NMS.

[jkjung-avt]

If your custom model is detecting more classes of objects, it would be slower.

Otherwise, I'm not able to make guesses with the limited information you've provided.

[xmuszq]

Hi Jk,

Thank you for your reply!
I trained in TF1.12.0 with the your suggested detection API version. My Jetpack 4.2.2 and tensorrt version 5.0.6.3.
I have only 2 classes.
What i have tried is to lower down the
topK=100,
keepTopK=100,
to 20 and increase the
'min_size': 0.05,
to 0.2
the fps can go up to 20.
But these settings will lower the performance.