ECCV 2016 Daily - Thursday

We spoke with Armin Mustafa , who is a Postdoctorate student to Professor Adrian Hilton at the University of Surrey . Her worked focuses on 4d dynamic scene reconstruction and has performed non-sequential alignment of partial reconstructions from non-rigid shape for single or multiple views. She works with Hansung Kim and Adrian Hilton , who is also attending the event. The main novelty is that existing approaches either require full reconstruction of the object surfaces. It also performs sequential alignment which is prone to errors if it drifts and fails to do rapid, complex motion. We are handling this by introducing a non-sequential alignment for partial surfaces from multiple views or single views. The main difficulty is that the scenes are dynamic, and the object is moving very rapidly. For example, it can prove very challenging if the legs of the object move or dance very quickly. In that case it would be very difficult to simply use optical flow. That’s why Mustafa and her team want to perform non- sequential alignment. With a large sequence of 1,000 frames, starting from frame 1 and going on to frame 1,000, there will be errors. Let’s say they align it and perform non-sequential alignment with a root node. If you go from one root of a tree to another, there is a shorter path length through which they can align the sequence very easily compared to a very long sequence of 1,000 frames. 4D Match Trees for Non-rigid Surface Alignment 12 ECCV Daily : Thursday Presentations