Abstract: Kinematic character controllers can learn to generate very realistic motions, but they generally lack flexibility. Physics-based characters can realistically respond to environmental variations and interactions, however, they could appear unnatural when performing unusual movements to maintain balance. I will talk about some recent works that use deep reinforcement learning to combine them to generate both natural and flexible motions.
Title: Towards markerless tracking of the honeybee waggle dance
Abstract: Honeybees perform a dance to communicate the location of valuable resources to nestmates inside the hive via a mixture of body motion, mechanical and chemical cues. Today, more than 70 years after its discovery much of the underlying biology is still not yet understood. I am interested in utilising a computational approach to explore possible neural circuitry underlying the communication. Towards this, analysing dance activity from raw video footage recorded within the hive could prove invaluable for testing some initial hypotheses. However, such data characteristically exhibits uneven illumination within or across frames, multiple targets, possible occlusions as well as noisy backgrounds. In this talk, I would like to introduce the challenge of tracking bees within the hive, different ways this task has been approached in the past and the possible benefit that a ‘self-supervised’ deep learning approach could offer.