Sensor: Amund Vedal, NR
Internal sensor: Anis Yazidi, UiO
Supervisors: Marius Aasan, UiO; and Adín Ramírez Rivera, UiO
Abstract: Video action recognition is an active field of research. Previous models used in action recognition have achieved impressive results in accuracy. However, video data is computationally expensive and these models often require large amounts of resources in order to be trained. This thesis explores methods for low computational cost in video action recognition. We extended a graph-based superpixel algorithm to a supervoxel algorithm and combined it with a graph attention network. We test two different versions of this algorithm, spatio-temporal connections and spatial first and spatio-temporal connections later. While our model does not reach the accuracy results of other state-of-the-art models, it does show impressive efficiency with only 1.2 GFLOPs for our third best model and 14.5 GFLOPs for our best-performing model. This indicates that using pure graph based models can be a method for highly efficient video action recognition.