Bigger drop in Detection Accuracy while running yolov7 standalone and in Deep Stream

[tkhan3]

I am trying to do person detection using yolov7. I ran my stream though yolov7 and it detected very well. Then I added yolov7 model to deepstream with procedure specified in (this) DeepStream-Yolo repo. Please note I am using full FP32 Engine.

I am seeing a very big drop in person detection accuracy. Any thoughts why its happening and how to address that ?

[cgrtrifork]

This is probably related to #339