Pixel Club: Approximate Nearest Neighbor Search for Video

Speaker:
Nir Ben Zrihem (EE, Technion)
Date:
Tuesday, 8.4.2014, 11:30
Place:
EE Meyer Building 1061

Is it possible to perform BM3D in real-time? Probably not, but it can be approximated. In this talk we will present an algorithm for video patch-matching that enables real-time video processing for a variety of applications, such as colorization, denoising, or artistic effects. We name our algorithm RIANN - Ring Intersection Approximate Nearest Neighbor - since it finds potential matches by intersecting rings around key points in appearance space. RIANN's real-time performance is attributed to two key properties: (i) RIANN leverages statistical properties of videos and adapts its search complexity to the amount of temporal difference between frames. (ii) RIANN employs a relatively long pre-processing stage, however, since this stage is global for all videos and all frames it pays off. We show via experiments that RIANN is up to two orders of magnitude faster than previous patch-matching methods and is the only solution that operates in real-time

Back to the index of events