No announcement yet.

TensorFlow/Vuforia performance

  • Filter
  • Time
  • Show
Clear All
new posts

  • TensorFlow/Vuforia performance

    Anybody have issues with TensorFlow or Vuforia performance? For some reason with our setup, when we activated TensorFlow, our main robot loop time degraded from less than 50 msec to more than 300 msec. This makes PID control almost impossible. We currently worked around this issue by activating TensorFlow at the beginning of the competition. Once we recognized the gold mineral, we killed TensorFlow, not just deactivating it, but calling "shutdown". Deactivating doesn't fix the performance problem but shutting down does. We tried many things to diagnose the issue, it seems our main robot thread is being starved when TensorFlow /Vuforia is activated. In fact, we even tried creating our own thread and it got starved too. Anybody have seen this?

  • #2
    BTW, we have Motorola G3 phones. They are not the fastest in the world but I believe they should be adequate. Even if TensorFlow/Vuforia may impact performance, I was expecting it has its own thread and will just not give me detection update until hundreds of msec later (i.e. has poor performance on its own thread) but hogging CPU from other threads is a little hard to understand.


    • #3
      I am going to take a look at this later today. We have Motorola G2 phones, so hopefully that won't cause a problem.


      • #4
        Sorry to respond so late on an old thread. But looking at my kid's loop times with tensor flow and vuforia running in iterative opmode they see loop times in the 80-100ms range with the usual occasional spikes higher with something in the OS garbage collects or whatever. Motorola E4 and current SDK and all that