pierre garrigues


I do research and development at IQ Engines where our goal is to extract relevant information from images. Our technology uses a combination of computer vision and crowdsourcing which makes it possible to attach keywords to any image, and to continuously improve our computer vision algorithms based on crowd responses. You can check it out by using our image recognition API or downloading oMoby on your iPhone.

I obtained my PhD in May 2009 from UC Berkeley in the Electrical Engineering and Computer Sciences department. I worked under the supervision of Prof. Bruno Olshausen in the Redwood Center for Theoretical Neuroscience, and also collaborated with Prof. Laurent El Ghaoui. I did my undergraduate studies at the Ecole Polytechnique in France.


pierre.garrigues at gmail.com




natural image statistics

It has been observed that natural images, as complex and varied as they appear, have an underlying structure that is sparse. That is, one can describe a given image with only a few features. Exploiting sparsity in natural images is at the heart of many successful image processing algorithms such as denoising or compression (I gave a guest lecture on the subject, check it out here). I am extending the sparse coding model to account for further structure in natural images such as the statistical dependencies among the features. Exploiting these leads to a richer model of natural images that can then be used in image processing and computer vision applications.

P. Garrigues and B. Olshausen. Group Sparse Coding with a Laplacian Scale Mixture Prior, to appear in Advances in Neural Information Processing Systems 23 (NIPS 2010). [pdf][supplemental][bibtex]

P. Garrigues and B. Olshausen. Learning Horizontal Connections in a Sparse Coding Model of Natural Images, Advances in Neural Information Processing Systems 20 (NIPS 2007). [pdf]

P. Garrigues and B. Olshausen. Learning Horizontal Connections from the Statistics of Natural Images, in Computational and Systems Neuroscience 2007, Salt Lake City, UT.


The operation to compute the sparse set of features that represent a given signal, for example an image, is the solution of an optimization problem commonly referred to as Basis Pursuit or Lasso. I have developed a fast algorithm to solve this problem when the observations are sequential, such that a path is computed from the current solution to the solution after observing a new data point.

P. Garrigues and L. El Ghaoui. An Homotopy Algorithm for the Lasso with Online Observations, to appear in Advances in Neural Information Processing Systems 21 (NIPS 2008). [pdf][poster][code]

video processing

We propose a new scheme for position coding of the atoms within the Matching Pursuit algorithm as applied to the displaced frame difference in a hybrid video encoder.

P. Garrigues and A. Zakhor. Atom Position Coding in a Matching Pursuits Based Video Coder, in VLBV 05, Lecture Notes in Computer Science 3893 / 2006, pp. 153-160. [pdf]


P. Garrigues. Sparse Coding Models of Natural Images: Algorithms for Efficient Inference and Learning of Higher-Order Structure, in Technical Report No. UCB/EECS-2009-71, May 20, 2009.[pdf][video]

P. Garrigues. Atom Position Coding in a Matching Pursuits Based Video Coder, Master's thesis, 2005 [pdf]