RQT plugins for Tensorflow image recognition

Contains rqt plugins for training a neural network for object recognition using TensorFlow.

Train plugin

This RQT plugin (http://wiki.ros.org/rqt) enables training a neural network to classify objects based on images using TensorFlow. The end result will be a trained TensorFlow graph and output labels as in https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/image_retraining/retrain.py

How to

Prerequisites

Annotated data with label subdirs that contain image segments of the specified label, e.g. /tmp/object_recognition:

/tmp/object_recognition/:
drwxrwxr-x 2 rein rein 4,0K okt 18 21:22 apple
drwxrwxr-x 2 rein rein 4,0K okt 18 21:22 banana
drwxrwxr-x 2 rein rein 4,0K okt 18 21:22 choco_peanuts

/tmp/object_recognition/apple:
-rw-rw-r-- 1 rein rein 14K okt 18 21:22 2016-10-18-21-22-29.jpg
...

/tmp/object_recognition/banana:
-rw-rw-r-- 1 rein rein 16K okt 18 21:22 2016-10-18-21-22-37.jpg
...

/tmp/object_recognition/choco_peanuts:
-rw-rw-r-- 1 rein rein 8,6K okt 18 21:22 2016-10-18-21-22-43.jpg
...

This labeled data set can be created with use of the label RQT plugin in the image_recognition_rqt package. But, feel free to use any kind of tooling to create your own data set!

Training

Start the train GUI

rosrun image_recognition_tensorflow_rqt train_gui

Set the images and output directory. For the example above, use /tmp/object_recognition as images directory

and for example /tmp as output directory.

Click the gear wheel in the right top of the GUI to set the batch size and the number of steps:

Once everything is configured, press the "Train" button.

Wait for some hours :) -- Check console log.

Could not convert RST to MD: No such file or directory - pandoc