Daily CVPR - Wednesday

Presentations 13 CVPR Daily: Wednesday Khurram: The practical applications are huge. For example, this can be used in surveillance tasks. Now you have millions of CCTV cameras all over the world. However, there are very few methods which can detect in an online manner. Let’s say I’m viewing footage from a CCTV camera, can I say what’s going to happen as I’m viewing it? That’s one aspect of it. The second aspect is in human computer interaction. Let’s say you are interacting with a computer in a Playstation or an Xbox. Now the action that is being performed in that interaction depends on what you do. This interaction needs to happen in a live manner. This can be used so you can estimate what the computer is doing, and the computer can know what you’re doing. CVPR Daily: What is the next step? Khurram: The next step is to find a real-life application and make it work so that people can use it. It was quite interesting and challenging when I started to work on this because this is a very new problem. People are not doing it in an online manner which is very necessary for the application. This gives a huge area that people can start working on. One interesting analysis that we can come up with using our method is that since you are predicting what the action is, we can say how much of the video you need to watch if you are looking for a certain action. If I’m looking for a certain action, let’s say kicking a ball, how much of the video do I need to see to be able to predict what the action is? We do the sort of analysis that can help and distinguish which actions are more challenging than the other ones. “ since we are predicting what the action is, we can say how much of the video you need to watch if you are looking for a certain action ”