Lower detection rate on small objects after conversion to TensorRt

[mchi]

please take a look Pytorch model deployed on deepstream no predictions - #6 by pyml to make sure you set the gie key infer properities correctly.

Thanks!