Origin of Ferroelectricity in Thin Film HfO2 Probed by Revolving STEM and PACBED

Abstract

This paper present a form to speed up the recognition of a video using information of the motion. The motion field is extracted from the compressed information available in the video JPEG format. They show that this information is similar to the obtained from Lucas-Kanade optical flow but is faster to extract. A video is split using a grids of different size then a certain number of cells is obtained from the video. Each cell is described using 4 kinds of descriptors HOG,HOF,MBHx,MBHy. Each kind of descriptors are quantized in a histogram (k-means) considering 8 bins of orientation and l1-normalized. To classify they use SVM multichannel. Also, the authors extract fisher vector from the video and classify then using SVM. The approach is tested using HOHA, UCF50, HMSB1 and UT-interaction.