In this talk, we present a novel algorithm for low rank LU decomposition that uses random projections type techniques and can be used to efficiently compute a low rank approximation of large matrices. The randomized LU algorithm can be parallelized and efficiently run on GPUs, making it suitable for image processing algorithms.

 

Moreover, it can utilize sparsity to accelerate the computations even farther. Several error bounds for the algorithm's approximations are proved using random Gaussian and random sparse matrices .

 

 

 

 

What is an image? There is a growing trend in the last decade or so to treat images as a bag of patches. This can be seen in texture synthesis, object and action recognition and even image denoising. This approach has great success but it comes at a price.

 

All geometric information is lost. In this talk I argue that much geometric information is implicitly encoded in the bag of patches, demonstrate how to recover it and show a number of potential image editing applications. Then, I will consider the problem of organizing images in the correct temporal order. This problem occurs naturally in group photography where a group of people take multiple images of a dynamic event.

 

Since the cameras are not necessarily synchronized we need to organize the images in their correct temporal order, a problem we term Photo Sequencing. Common to both cases is the need to establish what is the correct way to organize the visual data and then devising ways to achieve this order.

Joint work with Tali Dekel, Yael Moses, Taeg-Sang Cho, Moshe Butman and Bill Freeman.

 

 

 

The aim of SceneNet is to use the power of crowd sourcing, in the form of multiple mobile phone users, to create a higher quality 3D video scene experience that can be shared via social networks. A typical SceneNet scenario can be a rock concert, a sports event, a private or a public ceremony, breaking news events and any other multiple mobile users' crowded event.

The SceneNet pipeline starts at the mobile device where the video streams are acquired, pre-processed and transmitted along with a tagging identity to the server. At the server, the various video streams are registered and synchronized and then submitted to 3D reconstruction to generate a multi-view video scene that can be edited and shared by the users.

In this talk we will give an overview of the SceneNet project and the technologies we develop in order to facilitate it, from mobile computing to 3D reconstruction and GPU computing.

This project is done in collaboration between SagivTech, Ecole Polytechnique de Lausanne (EPFL) and the University of Bremen.
This project is funded by the European Union under the 7th Research Framework, programme FET-Open SME, Grant agreement no. 309169.