What's meanings of outputs in detectnet_v2?

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hi, I trained my own model based on Traffic Cam Net(Based on detectnet_v2）and I tried to deploy it on jetson nano, but I got these outputs.
boxes:[(826,262,458,-66)]
class:[14]
confs:[35.53800964355469]
I was confused that how can boxes have a negative value? And why confidence can bigger than 1?Can some one explain these data for me?
By the way, the code I used in jetson nano was made for SSD_mobilenet, and I have to make it work with detectnet_v2. So what’s the difference between SSD and detectnet’s outputs?

2

In tlt-converter command, there is an argument ‘-o’, for SSD_mobilenet it’s NMS and for detectnet_v2 it’s ‘output_bbox/BiasAdd,output_cov/Sigmoid’. Is the problem caused by difference of them?

4

Firstly , please check your postprocessing is correct.

For postprocessing of Traffic Cam Net(Based on detectnet_v2, please refer to /opt/nvidia/deepstream/deepstream/sources/libs/nvdsinfer_customparser/nvdsinfer_custombboxparser.cpp

More reference: Run PeopleNet with tensorrt - #3 by steventel

For the difference of model output, please refer to

In TLT the preprocessing of SSD image is like below:

assume RGB input values which range from 0.0 to 255.0 as float
change from RGB to BGR
then subtract channels of input values by 103.939;116.779;123.68 separately for BGR channels.

5

Thanks for your answer! I’ll try it right away.