Section: New Results

Designing criteria

• Criterion design and optimization methods for computer vision On the topic of large-scale image segmentation with multiple object detection, targetting as an application the analysis of high-resolution multispectral satellite images covering the Earth, challenges are numerous: scalable complexity, finding good features to distinguish objects, designing shape statistics as well as an optimization method able to incorporate them. We propose a solution [42] , [43] based on the construction of binary partition trees and on their optimization, whose cost is alleviated thanks to theoretical results reducing the search space. Concerning video segmentation, we have extended previous work, on the inclusion of shape growth constraints into classical MRF settings (graph cuts with globally optimal segmentation), to the case of multimodal sequences of medical 3D scans [19] . We also studied a new family of metrics in [9] , together with a redefinition of the associated gradient and practical ways to compute it. This allows the consideration of new types of priors on planar curve evolution, such as piecewise-rigid motions. Surprisingly, the problem of finding the best piecewise-rigid approximation of a motion turns out to be convex, and to be linked to sparsity approaches.

• Algorithm selection and configuration Two PhD theses are still related to the former Crossing the Chasm SIG: Nacim Belkhir has worked on inline parameter tuning for the CMA-ES algorithm in the context of a large number of cores [21] , and is now using surrogate models to compute the features of expensive continuous optimization (submitted). François Gonard's PhD is dedicated to algorithm selection. The original application domain is that of expensive car industry simulations (within the IRT-ROM project). Initial results concern combinatorial optimization, and François obtained a "Honorable mention from the jury" for his submission to the ICON Challenge (http://iconchallenge.insight-centre.org/ ), for its original approach coupling a pre-scheduler and an algorithm selector. A paper describing the algorithms and analyzing the results has been submitted.

• A statistical physics perspective In the topic of MRF design, with motivating applications in large scale inference problems like traffic congestions, we have finalized in [13] an approach based on the disordered Ising model relying on approximate solutions to the Inverse Ising problem. To this specific problem we also propose new approximate solutions, compliant with the generalized belief-propagation algorithm in the static [63] and a new $l_0$ regularized method based on a maximum likelihood maximization for the dynamical case[11] . In fact in [63] we have developped a method adapted to the generalized belief propagation framework, aiming at adressing directly and systematically the loop corrections without loss of scalability, offering new possibilities in the context of inference by MRF models. In parallel, a better understanding of the so-called mean-field approximation when the phase space is clustered has been derived [68] giving a direct method to solve static inverse problem in the weak coupling limit. Apart from the method point of view, some consideration over what can be said on the data has been considered, still in the topic of MRF design. In this sense, it is shown in [69] that the reconstruction of the MRF model depends strongly on how the data are gathered, and how to remove redundant data and keep a good reconstruction.

• Multi-objective AI Planning This activity had almost stopped since the end of the DESCARWIN ANR project. However, a productive intership resulted in some new benchmarks in the ZenoTravel domain together with an exact solver ensuring the knowledge of the true Pareto front [48] , [47] .