shape_learning is a library for simultaneously learning multiple shape models based on user input, such as demonstrations.

Initial models for over 100 symbols (mostly letters) are provided.

- matplotlib 
- scipy 
- recordtype 
- sklearn

The ShapeModeler class manages a model of a particular shape, for example a letter. The current implementation uses a principle component analysis-based model: it takes a dataset of instances of the shape and determines the parameters which explain the majority of variance in the dataset.

A GUI to visualise the effect of varying parameters of the shape model is provided in scripts/shape_model_gui.py. Sample usage:

./scripts/shape_model_gui.py 'd' 5

An example of the GUI for visualising the effects of the model of 'd'.

A set of principle components parameters for a set of shapes can be conveniently generated and saved with:

./scripts/generate_custom_parametrization.py <regexp to select shapes in dataset>

The ShapeLearner class manages the learning of the parameters of a particular ShapeModeler attribute. The parameters of user-demonstrated shapes may be determined from the model, and used to update the learned parameters of the system.

The ShapeLearnerManager class manages collections of multiple ShapeLearners. For example, in the context of learning words, the ShapeLearnerManager keeps track of the current word being learnt and its associated ShapeLearners, in addition to the information on each letter/word which has been previously seen. This allows for long-term system memory, even when the shapes are not always part of the active collection.

A sample application of learning words is provided in scripts/learning_letters.py (additional dependency of Kivy). The default dataset of letter instances is from the UJI pen charaters 2 dataset. Sample usage:

./scripts/learning_letters.py 'case'

An example of the app for demonstrating the letter 'a' (left) to update the system-learned shape (right, originally 'o'-shaped).

For a more complex use case of the shape_learning library, see the CoWriter project.

The current dataset of shape (or a subset thereof) can be vizualised with:

./scripts/vizualise_dataset.py <regex>

For instance:

./scripts/vizualise_dataset.py "[a-z]"

A set of predefined principle components values can be provided (with --parameters) and the resulting shapes while be overlaid one the default mean shapes:

Scripts are provided to generate the datasets for letters (and many other symbols like numbers) from the UJI Pen Char 2 database.

Run:

python ./tools/dataset-preprocessing/parseDataset.py <path to ujipenchar2.txt> <dest path>

to parse, interpolate, normalize, and cluster the letters present in the database.