Over the Easter weekend I finally got around to implementing a first prototype of an idea I’ve had for a while, which aims to bring some state of the art computer vision techniques to mobile devices.
Deep Vision uses the implementation of convolutional neural networks provided by libccv to classify images. So it’ll try to figure out whatever is the principal object in an image your provide it with.
At the moment it just has a sample classification database from the ImageNet project, containing 1000 assorted items, however in the future I’d like to see specific classifiers for different tasks (e.g. a classifier trained purely on different plants, so when you’re out for a hike and you want to know what something is you can just point your phone at it and find out.)
Unlike something like Google Goggles it’s doing all the classification on the phone itself without needing to upload the image to any external services.
The video below provides a quick demo of it in action and you can also grab a click package here to play with it yourself: http://mikeasoft.com/~mike/com.mikeasoft.deepvision_0.1.3_armhf.click
Source code can be found at: https://launchpad.net/deepvision
It was just hacked together over the weekend, so it’s still a little rough in places but all the core functionality should work reasonably well :).