Low accuracy nvinfer + nvtracker

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU): Jetson Nano
• DeepStream Version: 5.0 GA
• JetPack Version (valid for Jetson only): 4.4

Hello everyone, I am writing this topic because I have a series of doubts regarding the results I am having with a deepstream application. The application in which I work contemplates the use of nvinfer + nvtracker + nvdsanalytics. The video source is IP cameras (RTSP).

The configuration of the cameras is as follows:

  • Codec H.265
  • FPS: 21
  • Bit Rate: 2048
  • Bit Rate Type: VBR
  • Resolution: 1920x1080

The deepstream application in general is configured as follows:

  • Model: resnet10 (resnet10.caffemodel_b2_gpu0_fp16.engine)
  • Pgie interval: 4
  • streammux: 1920x1080, livesource = 1
  • nvtracker: NvDCF with default configuration file

Here you can see the application working on a demo video file:
cemzoo_output_ok

Here instead, it is seen in real operation in an apparently less complex video sequence:
cemzoo_output_bad_2
*If someone wants to see the video file at higher resolution, they can request it

I do not know if I am correct, but my impression is that there is a decrease in the accuracy of the detection. Also I think that the malfunction of nvinfer causes a malfunction in nvtracker, since the detected object is lost quickly. This raises the following doubts:

  • Do you think it is a problem with the location of the camera? If so, is there a document with recommendations in this regard? (distances, angles, etc)
  • Do you have information regarding the images with which the resnet10 model was trained? for peoplenet there is information about it.
  • Are there nvinfer or nvtracker configuration parameters that must be adjusted depending on the scene and allow to improve this performance?
  • Is there a problem that I am not noticing?

Check Dim size in .prototext files and NVIDIA Metropolis Documentation

Hello @bcao , I appreciate your answer, I will do tests by changing the prototxt values. Thanks also for the link to the thoubleshooting page.

I will comment on the results when I have them.