Computer Vision News - April 2016

Application COMPUTER VISION NEWS Giving up frames and concentrating on objects involves challenges which are more conceptual than technological: on one hand, it was obviously difficult to ask people to give up chronology in their videos; on the other hand, it was not even clear that this direction would offer the expected results. Once the hard decision is taken, the next task is to track the objects. This need to be much more precise than in other projects, since time coordinates are missing: if object segmentation is not performed with the highest precision, the system might separate object parts and show them disjointedly: a head detached from its body and the like. Segmentation and tracking requirements are therefore more severe than they regularly are. Deep learning and neural networks technology was crucial to achieve successful classification and segmentation tasks. But the decision to observe and decide what is interesting and how to act is still human. The next step was to select interesting objects: only cars or only bicycles and so on. That brought to the addition of filters : parameters can be defined to select only people, or identify any information we are looking for. That makes the video summary even more efficient, since only relevant items are displayed. It makes it also easier to recognize objects behaving differently than others: when all bikers pass at similar speed, we can instantly recognize the one riding much faster. Anomalies and irregularities are easier to identify when only the category they belong to is shown. Besides the obvious interest of this application in the fields of surveillance and security, Prof. Peleg sees future opportunities in completely different realms like the home market: if yesterday you couldn’t know at what time the gardener arrived, now you can find out in seconds! 11 “ Going beyond chronology breaks the biggest of video taboos ”