Inside/Outside project

We built the Inside/Outside Project for the Create Together Day at the 2017 Citizen Science Conference. The goal of the Inside/Outside Project is to train TensorFlow models that will classify whether any particular photo was taken inside or outside. The filtered dataset can then be provided to a second human-aided classification step; for example, to help build a repository of public places.

The Inside/Outside project involves three components:

1. This repository, which contains the wq-powered client application and observation database for collecting training images.
2. The TensorFlow+wq model broker database, which passes the classified training images on to the retrainer
3. The CitSci2017 TensorflowRetrainer, which retrains Inception on the training data and uploads the resulting model back to the broker.

Implementation Details

This application was created with the wq start tool. The revision history for the initial version documents the full process:

1. ad70796 Initialize with wq start 1.0.0rc1
2. 274dfc1 Configure SSL with LetsEncrypt
3. 8f89275 Add XLSForms for Category and Observation
4. cab2327 Enable wq/locate.js
5. Customize workflow:
   • 3f244b4 Use <input type=tel> rather than <input type=number> to get around precision issues with lat/long
   • 25e263b Don't require authentication to submit a photo
   • ea70810 Customize category screen (fix pluralization and ordering)
   • f3acccd Customize the links on the home screen
   • a0cd5d2 Set defaults for date and location mode on observation screen
6. c7a6fac Integate the cordova TensorFlow Plugin