CVPR Daily - Tuesday

It uses a memory-based algorithm, following the recent idea of memory- augmented networks in a very smart way. Typically, backpropagation is very slow . If you want to backpropagate your network at runtime, you have to wait too much time. The idea is to use a memory which is very fast to be loaded and unloaded at runtime. It’s a challenge, because the system is running in real time. The real-world application of this is open world face recognition . Federico explains: “ Typically, with face recognition you have the bad guys already labelled. In this approach, anyone can in principle be there. If I don’t know you, you go into a gallery. It’s a dynamic gallery, so the gallery of the bad guys is not fixed and is built at runtime. It’s an open border .” In terms of next steps, Federico says that is about learning the representation in real time. In this approach, the representation – which is a descriptor taken from the VGG (Visual Geometry Group) Face , a huge deep neural network pre-trained at face recognition – is fixed. The idea is that once you have collected so many faces, you can fine-tune on top of these new faces. You are all invited to learn more by coming along to Federico’s poster [Q9] today, at 12:30-2:50 in Halls C-E. Tuesday 9 Federico Pernici In video, you are always viewing the same person with the same appearance, so after a few seconds, you don’t need any more data on that person.