Due to the exponential growth of video content over the last decade, there has been an increasing need for video analytics. Using video analytics has many practical benefits, such as helping in the monitoring of surveillance imagery or generating video previews on YouTube.
At summy, we are developing a video analytics system specially designed for TV content. Currently, we generate metadata based on a large number of explicit features that are generated by visual and audio networks. However, we are interested in the capabilities of an end-to-end approach, which would help to make our solutions less domain-specific.
A number of methods have been proposed for solving problems in this field, for example using Adversarial LSTMs or using a Reinforcement Learning approach. While these methods can serve as a good baseline, they often have room for improvement. We are specifically interested in performance on longer videos instead of the short videos in scientific datasets. We are looking for a student that wants to tackle these problems in a master thesis.