Something wrong with the secondary engine do infer with its batch size

Hello,

I am running a deepstream-app using two engine detector (base on ssd); The first one is detected traffic (motorbike, car,…) the the second one is detected license plate. I put one mp4 video file, the app run fine but with a vehicle was detected, it have many license plate show on it (not same location; event it show out of the parent’s bounding box area??); although I set keep top k of the secondary (license plate detector) is 1 with confident is 0.7. Cause I mean each vehicle should have the only one its license plate.

I checked it out why have such strange problem. And find out that the secondary should have the batch size is 1 not like I set to 4 before.

But the point here is that when have multiple source input the batch-size is not 1 anymore. And I imaged that each object vehicle was detect by the primary engine was forward to the secondary engine one times; but maybe somehow with batch size of secondary engine is 4 it put the object vehicle to do infer again? Maybe it is the reason why each it have many license plate show on it like I said.

Please check it out, and tell me how I can do now?

Hi,

May I know how do you implement the dual detector pipeline.
Do you follow our back-to-back-detectors sample here?
https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/tree/master/back-to-back-detectors

Thanks.

Hi AastaLLL,

Yes I do follow back-to-back-detectors; but the model is different; it’s is ssd model with 3 labels (background; license plate and face)

Hi AastaLLL,

Sorry to bother, I find out that my purse boxes custom filter boxes wrong with batch size of the secondary engine; so I fixed it; Thank you for your reply;

Please closed the topic.

Good to know this.

Thanks for your update.