Abstract
We propose portrait retrieval from moving pictures with symbolic rough sketch, which is drawn frequently. The posture of target people is shown by symbolic rough sketch. And lines around the symbolic rough sketch show the movement of target people. In addition, we take some descriptors of MPEG-7 as metadata to calculate similarity between drawn query rough sketch and compared image in database. We also propose a discriminative conversion algorithm to make it possible that symbolic rough sketch drawn by users reflects their intention more precisely. Experimental results on an indigested home video, one of the practical applications of portrait retrieval, show that our algorithm improves the value of similarity between rough sketch and user's targeted image and that we can retrieve our target portrait with symbolic rough sketch.
