Ming Zong

Citation

Thesis Title
Deep Learning for Action Recognition in Videos

Read article at Massey Research Online:

Action recognition is an important research topic in computer vision and has been successfully applied to many vision-related practical applications, such as human-computer interaction and smart video surveillance. However, the accuracy of computerised action recognition is not commensurate with the human response. Mr Zong proposed to directly perform 3D convolutions on a multiple cue-based input, instead of a single cue-based input, for action recognition. He demonstrated that directly performing 3D convolutions on multiple cues achieved higher accuracy than a single cue. He proposed a motion saliency stream to capture the salient motion information, and a spatial saliency stream to capture the spatial saliency information. Experimental results showed that both spatial saliency information and motion saliency information is helpful for improving the classification accuracy.

Supervisors
Professor Ruili Wang
Dr Andrew Gilman
Professor Xun Wang