GroundTruth

Tool for annotation of images with ground truth that can later be used for machine learning. The idea is to improve the performance of an existing detector by increasing the amount of training data used. The detector is executed on some new unlabeled data and the result uploaded to a server. The app is then used to correct any mistakes made by the detector and thereby form a new set of labeled training data.

The app will ask a question, such as "Are these faces?" and show a set of images that the detector believed was faces. The user is supposed to cross out any images he is not sure belongs to that category by tapping them. Make sure to read the questing carefully to cross out the correct images. Sometimes it's the majority of the images that should be crossed out. Also cross out any images for which the answer is unclear. Crossing out an images does not automatically place it in the opposite class. It will however make it show up in a question about the opposite class later on.

The app is based on the http://kivy.org framework and uses icons from http://www.icons-land.com