Prototype for non-verbal access to information, targeting users with intellectual disability. This prototype shows a UI which allows browsing a large database of images. At each stage, the demo shows a target image at the center and thumbnails of images similar to that image around it. There are several combinations of image selection and sorting algorithms:
- closest in term of similarity
- diverse using an MMR-like algorithm
- 1d sorting, top-left is most similar
- 2d sorting, close-together images are most similar
The algorithm uses image representations generated with a deep convolutional neural network such as ResNet.
To run, make sure you have python3 accessible as python, virtualenv and gcc installed.
./run.sh
This prototype can accomodate for multiple data sources. Each data directory shall conain:
images/
a directory with thumbnailsnames.txt
a file containing am image thumbanil filename and a source image path separated by a tab on each linevectors.npy
a numpy matrix of shape number of images times dimension of representation where each row corresponds to the image on the same row innames.txt
Representations are typically low-dimension dense vectors (for example size 256) and can be generated with CNNs such as ResNet.