05/05

Found and fixed an issue in the integrated classifier that was causing some of the extra time the added to the latency for detecting images!!  Sadly this was after the demo.. but now we surpass all our requirements.  Time to detect gesture has gone down from ~7 seconds to 1-2 seconds :))  We will write about this in the report and hope the professors see.

The problem was that we were opening, reading, closing the saved model files and then loading and compiling the model and the weights every single time a frame was passed through the classifier.  Changed this to open, read close model file, and then load and compile model only once in a setup_classifier function.  For some reason could not do loading weights only once because it caused classifier to hang.. unknown why.  Total time saved = 5 seconds per prediction.  YAY!

Adding a neural stick would probably still benefit and make this close to real time.