|Description (include details on usage, files and paper references)||The VidPairs dataset contains 133 pairs of images, taken from 1080p HD (~2 megapixel) official movie trailers. Each pair consists of images of the same scene with usually some motion of both camera and subjects in the scene (The images are between 1 and ~30 frames apart in the video). We note that pairs of images with only slight camera and subject motion are not very challenging in the dense patch matching framework and could be handled specifically via registration or optic flow techniques.
Coherency Sensitive Hashing [pdf]
Simon Korman, Shai Avidan
ICCV 2011, Barcelona