DetectNet Python not giving the same output

I’ve been trying to use detectnet on a TX2 for people detection but the output of the network seems inconsistent. When run through the jetson-inference sample code the network outputs much more accurate labels but when run run using python and caffe there’s barely any detections at all.

I’m running the current version of NVCaffe, Jetpack 3.0 and using the inference sample code from https://gist.github.com/lukeyeager/777087991419d98700054cade2f755e6. I’m running the multiped-500 reference model but I think the same issues occur on others also.

Perhaps it’s the thresholds and parameters used by the clustering layer? Otherwise the networks should be the same. jetson-inference has a simple C++ clustering algo I wrote for ease-of-use which seems to work pretty well in most situations, however you can tweak the parameters (same with the Python layer, except it has more parameters).