Abstract
Probabilistic inverse graphics has the potential to revolutionize
visual inference in the face of uncertainty. However, the prerequisite knowledge to
develop state of the art systems is immense. In this work, we simplify the issue
and target optical digit recognition, one of the simplest inverse graphics problems
known. While the task itself has largely been solved, we aim to provide
a intuitive gateway into the world of probabilistic inverse graphics
to ease development of more complex systems.
To this end, we present a system to recognize and reconstruct
the ten basic digits by modeling each as a collection of 2D Gaussians.
Running inference on noisy point set images, we demonstrate a success
rate of 0.95 over 50 trials per digit. Furthermore, we present
two of our past attempts at this problem and provide intuitive reasoning on
their strengths and
pitfalls