Technology Forum Has Moved!

The FIRST Tech Challenge Technology forum has moved to a new location! Please take a look at our forum blog for links and instructions on how to access the new forum.

The official blog of the FIRST Tech Challenge - a STEM robotics programs for students grades 7-12.

Note that volunteers (except for FTA/WTA/CSA will still access their role specific forum through this site. The blog also outlines how to access the volunteer forums.
See more
See less

Duck not recognized in Tensor flow

  • Filter
  • Time
  • Show
Clear All
new posts

  • Duck not recognized in Tensor flow

    We set up a webcam (logitech 920) on our robot for autonomous use. We modified the sample op mode to do so and it works very well on our mat. We recently had a league meet at a local school and Tensor Flow would not recognize the duck, we could see it using the stream video on the drivers station but the program would not put a box around it and say its a duck. The overhead lighting in the gym seemed bright enough, much like at our mat we work on. We ended up setting the confidence level to .5 and it saw it one or two times in the remaining matches. The next day we put it on our mat and it saw the duck immediately and showed a .97 confidence. I turned out all of our lights except two over head fluorescent night lights that I cant turn off. It still saw the duck instantly. So, I asked the school what kind of lights they had in the gym where we competed. They had the lights redone and now have LED lights. Is there an issue with LED lighting that the tensor flow cant see the duck??? If so, how do we work around this. I found out our league tournament coming up has LED lights in their gym also.

  • #2
    Maybe try adding a non focused incandescent or halogen light source. Read the game manual carefully and Q&A forum to see how to legally power. Pay attention to controlled vs non controlled.


    • #3
      In a comparable LED lighting environment, see if these webcam controls help:

      Your testing should include white balance, newly added in SDK 7.1.

      These controls can be made available in Blocks, by creating myBlocks in OnBot Java:


      • #4
        The closest I am able to replicate the environment we were in is shielding the robot and duck on the mat with cardboard, so it doesn't get any of the fluorescent light directly on it, and it doesn't see the duck. I attached a Rev Robotics Blinkin with a strip of 12 volt RGB LED's to the robot. Same issue (probably because it's not a full spectrum strip, its individual red, green and blue LED's. If I shine the flash LED on my phone on the duck, it will see it, but those are a special type of LED that are designed to get the best color/exposure for pictures.

        I had some very bright white LED's that are used in commercial signs that are 12 volt. I connected three to one of the Rev motor controls and turned them on and off in Blocks (which is a legal way according to Game Manual 1). As long as I turn them on during INIT when Tensor Flow is initialized, it see's the duck almost instantaneously when we push play and also if you look at the camera stream. I don't know that there is any rule that you can't turn lighting on during INIT, but that is one thing I'm not sure of. If we wait and turn it on after hitting play, it takes much longer to recognize the duck.


        • #5
          Do you have all three markers in the field of view or are you moving the camera? We had a similar issue but ours turned out to be we were recognizing the markers and we had not differentiated in our code so it appeared that we were not recognizing the ducks but we’re recognizing too many things