Peoplenet inference shows strange output


I am trying to integrate the peoplenet pretrained model in a custom python application without deepstream. I put together a standalone application that takes an input image and loads the engine file (created using tlt-convert) and runs the inference.

I was able to setup preprocessing and post processing with inputs from several threads in this forum. However, I am getting outputs with classes >3 and with confidence values >1.0

Model fails to detect people in several test images we tried. What’s the correct way to interpret the network output? Please see attached python script.peoplenet.txt (5.2 KB) and output output.log (141.1 KB)

Steps To Reproduce

python3 -i input.png -o output.png