I changed the yolov3 network. The stride of the upsample layer is changed to 4 instead of 2. Using this model with deepstream trt_yolo_app, we observe a bounding box at the bottom of the frame for every prediction whereas the model works fine with darknet code. I suspect that this erroneous behavior is because of the incompatibility of the upsample plugin layer in deepstream yolo app with stride >2.