Measure the results quality such as MOTP MOTA id switches for nvtracker from DeeepStream 5.1

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

• Hardware Platform (Jetson / GPU) Jetson nano 4GB
• DeepStream Version 5.1
• JetPack Version (valid for Jetson only) 4.5.1
• Issue Type( questions, new requirements, bugs) question

hi,

My application is multiple person detection and tracking

I implemented the application using PeopleNet as below
Object-detection model: NVIDIA-developed DetectNet_v2
Backbone : ResNet34 as feature extractor
Model version: pruned_v2.1 - ResNet34 floating point deployment model

with the following
Object-Tracking: NvDCF

with pipeline:

filesrc → h264parse → nvv4l2decoder-> nvstreammux → nvinfer → nvtracker → nvvideoconvert → nvdsosd → nvegltransform → nveglglessink

Now i would like measure the results quality as follows:
– Multi-object tracking accuracy (MOTA),
– Multi-object tracking precision (MOTP),
– Identity switches (ID)

I have no idea how to do this in DeepStream or anyother way. Can someone pls give me clue?

3

We will check internally and update you ASAP.

4

Thank you, I eagerly look forward to your response!

5

User can check MOT challenge where there is a framework for this. Latest one allows bringing own dataset. User needs to enable kitti dump with tracking info in deepstream app and then convert kitti data to MOT data format.