Sauradip Nag

In this project, we aim to develop models for recognising complex and hierarchical visual activities. The current state-of-the-art methods such as variants of the two-steam network or 3D CNNs work very well on sports/film type videos, but struggle with complex activities with hierarchical structure such as cooking and assembling flat pack furniture. Analysing these activities is only achievable by delving deep into the interaction between humans and objects and compositionally modelling the hierarchical structure. To this end, we will develop novel compositional learning models and exploit deep graph convolutional networks for structured data analysis.