Jan-08-2019, 04:12 AM
There's nothing wrong with changing third party code. I just wanted to know.
I am curious about application though; partly because I'm not an expert with Tensorflow. Based on what I do know, it seems that you want to take a visual input from an onboard camera and evaluate obstacles in the way. Does this function weigh the objects it identifies based on distance? If not, then the robot may still collide with obstacles because its turns may occur at incorrect times. The problem being that a picture is 2D so basing 3d movement on it is tricky at best; it only works for us because we have two eyes.
The function evidently adds boxes to the image data. Would it be possible to evaluate the output image for those changes and base the turns on that?
I just noticed something about your class argument. The data passed in is:
I am curious about application though; partly because I'm not an expert with Tensorflow. Based on what I do know, it seems that you want to take a visual input from an onboard camera and evaluate obstacles in the way. Does this function weigh the objects it identifies based on distance? If not, then the robot may still collide with obstacles because its turns may occur at incorrect times. The problem being that a picture is 2D so basing 3d movement on it is tricky at best; it only works for us because we have two eyes.
The function evidently adds boxes to the image data. Would it be possible to evaluate the output image for those changes and base the turns on that?
I just noticed something about your class argument. The data passed in is:
np.squeeze(classes).astype(np.int32)To me, that indicates that all the data returned in that structure would be converted to 32-bit integers. Then, in the body of the code, you're checking if the class equals "turnRight", which is a string. Can you verify that the data returned is still a string?