Sorry, a king of dumb question:

I'm a bit confused about the purpose of the ObjectTracker node in the pipeline.

I understand that it will help track a defined set of objects from the mobilnetDN or mobilnetSDN (or yolo) parsing ImgDetections or SpatialImgDetections messages.

I run DepthAI integrated in a development environment: Touch Designer. Very useful as it natively provides all the tools to retrieve the streams (data, video, depth ...) and the complete API. You mostly just have to create the pipeline and it takes care of the rest. If anyone is looking for a tool to test and play with the dephAI API, I highly recommend it.

I ran this test to understand the workflow and simultaneously output the stream from the mobileetSDN. SpatialImgDetections and the stream from the objectTracker.tracklets.

Apart from being able to restrict the tracking to a certain type of object, I didn't notice any improvement in the tracking.

I mean, the spatialImgDetections and the tracklets give me about the same result. I haven't tested all object trackers yet.

Can you confirm my understanding that the objectTracker is supposed to improve the tracking of the Mobilnet NN or Yolo NN?

thanks

  • jakaskerl replied to this.

  • XvGt Can you confirm my understanding that the objectTracker is supposed to improve the tracking of the Mobilnet NN or Yolo NN?

    Not improve, but add tracking. MobileNetSSD and YOLO don't have tracking, they are running inference and return a detected object on the current frame. They don't care about previous or next frame.
    What object tracker does, is create an ID for each detection and then track that ID across multiple frames. This way you are able to check whether a certain object has passed a line, or implement people counting and such.

    Thanks,
    Jaka

      XvGt Can you confirm my understanding that the objectTracker is supposed to improve the tracking of the Mobilnet NN or Yolo NN?

      Not improve, but add tracking. MobileNetSSD and YOLO don't have tracking, they are running inference and return a detected object on the current frame. They don't care about previous or next frame.
      What object tracker does, is create an ID for each detection and then track that ID across multiple frames. This way you are able to check whether a certain object has passed a line, or implement people counting and such.

      Thanks,
      Jaka

        Hi Jaka ,

        Thanks for your clear explanation.

        I got it .