Cross-Angle Pose Search (Master Project, Finished)
Hand gestures are considered as one of the main non-verbal human communication channels and understanding and analysing them is an ongoing research in the field of cognitive science and linguistics. Despite the existence of a large amount of videos collected from TV footage, finding the instances of hand gestures requires hours of detailed screening of videos and manual annotations.
This project proposes an addition to the existing deepmime system to extend the hand gesture retrieval from different camera angles using skeletal data. The project aims at using the probabilistic pose embedding to project human poses from any angle into a unified space. This method will enable the deepmime search system in the first stage to retrieve similar poses and possibly hand gestures to the query video.
The project results will be evaluated on the basis of (a subset of) the News Scape video collection.
Start / End Dates
2021/04/01 - 2021/06/30