Team 2783 KYEOT Object Detection with Tensorflow

Also available on the Google Play Store: https://play.google.com/store/apps/details?id=coledev.kyeot.tensorflow

This app is a modified version of the official Tensorflow Object Detection app demo which can be found here: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/android

This has been modified to communicate with the RoboRIO based upon Team 254's code which can be found here: https://github.com/Team254/FRC-2017-Public

The most important part of this app is the frozen inference graph which can be found under the "assets" folder. This essentially is a set of weighted values that tell Tensorflow what a Power Up cube actually is, allowing it to be recognized.

The neural network was trained using a methodology that can be found here: https://github.com/CorvetteCole/Image-Net-Training

If you need help or have questions about the app feel free to contact me at corvettecole@gmail.com and I will try to help you

To do

• Drastically reduce latency by doing things in the Firmware development header
• Remove reference to AppContext in MultiBoxTracker (or at least reduce it). This is used far too often (every vision update) and shouldn't be. DONE
• Optimize inference graph for lower latency. This will involve generating a new inference graph from the raw output of the neural training and then quantizing it. DONE
• Confirm that timestamps are being applied to targets accurately. This is very important for latency compensation and I have no idea if the method I used to apply timestamps is anywhere near accurate. I hope it is but this is an important test. DONE

Firmware development

Install firmware on the Nexus 5's Hexagon DSP to accelerate neural processing drastically. See here: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/hvx

Use bazel instead of CMake as app build method and import new Tensorflow libraries built with Hexagon DSP support (stock library used by cmake does not)