MEDIABLENDER : INTERACTIVE MULTIMEDIA SEGMENTATION

This project aims at providing segmentation tools for assisting the browsing of multimedia content. We implemented several segmentation algorithms: two based on the Bayesian Information Criterion (BIC), the first frame-by-frame and the second using a 'divide- and-conquer' approach; a third based on the self-similarity matrix. We integrated the Torch3 library to benefit from classification algorithms. We also introduced optimization methods, notably parallelized image/video feature extraction distributed over GPUs and CPUs of one computer, using StarPU. All the components developed as part of this project have been incorporated in the MediaCycle framework.