A couple of questions regarding LPD/LPR

[yuweiw]

@foreverneilyoung

What makes the difference between the PGIE used model resnet18_trafficcamnet_pruned and the “original” resnet18-trafficcamnet used in all other samples? I have compared both w.r.t. GPU performance and I see a slight advantage for the pruned model, but what exactly has been pruned and why this “extra model”?

You can refer to How to use this model.

@foreverneilyoung

I was trying various variations of these models, with batch size 1 and/or fp16. While all models worked with B1, some spit warnings. My impression was, that fp16 degrees the performance parameters, while the inference results are better.

Some of it may be hardcode in the open source, you can look at the code logic like deepstream_lpr_app.c

@foreverneilyoung

what is the meaning of this warning?

It’s just warning message, you can ignore it.

@foreverneilyoung

I couldn’t manage to run a yolov4_tiny_usa_deployable.etlt_b1_gpu0_int8.engine. I could create the engine, but during runtime it “crashed” like so

You can try to set the parameters by referring to the FAQ.

@foreverneilyoung

Is it understandable, that an FP16 engine has higher performance costs compared to an INT8?

Yes.

@foreverneilyoung

And does it make sense at all to reduce the engine to the maximum number of expected inputs at all?

What do you exactly mean by that?

@foreverneilyoung

A third problem which I also observed with the original C app:

Could you attach how you confirm that? Thanks