Computational vision is one of the most challenging research domains in engineering sciences. The aim is to reproduce human visual perception through intelligent processing of visual data. CVC aims at proposing innovative techniques towards automatic structuring, interpretation and modeling of big (visual) data.
CVC was established on Sep. 2011 and is also associated with Inria Saclay, Ile-de-France unit through a joint project-team (GALEN). Its primary objective is to become a center of international scientiﬁc excellence in the ﬁeld of computer vision, machine learning and medical imaging analysis.
We have been pursuing a principled mathematical framework to formulate visual perception in terms of inference. In such a context, ﬁrst the solution of the desired vision task is expressed in the form of a parameterized mathematical model. This model is either data-driven, where annotated data sets and machine learning are used to determine an appropriate model space and constraints that capture the problem structure and observation patterns, or determined using real-world assumptions (what we often call physics-driven) through explicit natural modeling of the task to be addressed.
Given such a model, the next task consists in assessing the ﬁdelity of the model parameters to the available observations, namely establishing a model-to-data association. The aim of this task is to measure to what extent a given parameter choice complies with the visual observations and thereby drive the ﬁtting of the model parameters to visual observations. In simple words, this mounts to assessing how good a candidate solution is for expressing and ﬁtting the given data. This is achieved through the deﬁnition of an objective function on the parameter space of the model. Given the deﬁnition of the objective function, visual perception is addressed through its optimization with respect to the model parameters.
To summarize, computation visual perception involves three aspects, a task-speciﬁc deﬁnition of a parametric model, a data-speciﬁc association of this model with the available observations and last the optimization of the model parameters given the objective and the observations.
Such a chain processing is bound to have important shortcomings if developed separately, without a common underlying formalism. One of the main assets of CVC is that it involves experts in each part of the chain who are all well-versed in the mathematical framework of probabilistic graphical models. Using this common theoretical underpinning CVC aims at developing an integrated framework for training and optimizing such models by bringing in interdisciplinary techniques from the intersection of applied mathematics, computer science, statistics and engineering sciences, while leveraging on close ties with industry to develop innovative techniques for the automatic structuring, interpretation and modeling of big visual data.
The objective of our research is to overcome the aforementioned challenges in the context of a diverse set of computer vision and biomedical image analysis problems. Below, we list our objectives, which we group into three categories: (i) machine learning and optimization, which is concerned with the development of generic algorithms that are applicable to any graphical model and address fundamental problems transversal to artiﬁcial intelligence; (ii) computer vision, which is concerned with the adaptation of generic algorithms and the development of novel techniques to tackle speciﬁc problems concerning natural images; and (iii) biomedical image analysis, which is concerned with the adaptation of generic algorithms and the development of novel techniques to tackle speciﬁc problems concerning medical acquisitions. In the subsequent subsections, we provide the details of our contributions in the three categories.
- Machine Learning and Optimization. Exploitation and development of novel, eﬃcient of novel learning and inference techniques. In order to allow the use of inexpensive, large datasets, we focus on weakly supervised learning, where a large percentage of the training samples are only partially annotated. In order to prevent overﬁtting, we use principled and computationally feasible regularizers for the parameters of the model. We also focus on designing eﬃcient optimization techniques for performing inference, which requires the computation of the most likely assignment of values to the variables of the model for a given input.
- Computer Vision. Design of feature detection and description algorithms. Such algorithms are crucial to the success of computational vision perception algorithms. In particular, at this stage one can elliminate aspects of signal variation that are irrelevant to subsequent tasks, while at the same task elliciting those signal components that subsist further processing. This provides the subsequent stages with ’better behaved’ and more informative inputs, that facilitate accurate learning, and potentially also optimization in lower-dimensional spaces. In addition, we also develop eﬃcient inference algorithms for models that are of particular interest in computer vision, such as star-shaped models.
- Biomedical Image Analysis. Introduce a novel, modular (in terms of clinical applications, mathematical perception models and bio-imaging signals) framework. Exploiting the theoretical advances of our group in practical, high-impact applications is one of our main priorities. The dimensionality of the solution spaces encountered in medical imaging and the complexity of the high-order/recursive scoring functions require customized approaches to optimally exploit the task-speciﬁc problem structure.