Research Axes
Motion analysis and image segmentation
[+/-]
Data assimilation in geophysical images for motion estimation, curve tracking
|
The highly deformable nature of fluid flow, the complexity of the physic processes involved, but also the partially hidden measurements available from image data make difficult a direct
use of conventional image analysis techniques for tasks of motion estimation, detection, tracking and characterization. We face these issues using variational data assimilation tools. Such techniques enable to perform the
estimation of an unknown state function according to a given dynamical model and to noisy and incomplete
measurements.
|
See Project page (PI:COSTEL)
[+/-]
Phase field modeling for image segmentation
|
We address the issue of segmentation of entities that have the form of a ‘network’: the
region in the image corresponding to the entity is composed
of branches joining together at junctions, e.g. roads or vascular
networks. We introduce a new phase field higher-order
active contour (HOAC) prior model for network regions,
and apply it to the segmentation of road networks from
very high resolution satellite images. The model allows separate control of branch
width and branch curvature.
|
[+/-]
Turbulent motion analysis
|
We propose a new model to estimate and analyse turbulent fluid
motion from remote sensing image sequences. We define the fluid flow
by the scalar transport equation, and link the concentration of the scalar
to the observed grey-level in the image. We further apply a Large Eddy Simulation
(LES) decomposition and model the influence of
small scales with a subgrid turbulent viscosity factor.
|
Image understanding and learning
[+/-]
Towards optimal naive bayes nearest neighbor classifier
|
Naive Bayes Nearest Neighbor (NBNN) is a feature-based image classifier that achieves impressive degree of accuracy by exploiting ‘Image-to-
Class’ distances and by avoiding quantization of local image descriptors.
We improve its generalization ability by relaxing a too restrictive assumption and then dealing with parameter optimization using hinge loss minimization.
Our new formulation enables to compute optimal combinations of features and classification by detection.
|
|
[+/-]
Action recognition in videos
|
We represent a video sequence by a set of middle-level parts. A part, or component, has consistent
spatial structure and consistent motion. To further exploit
the interdependencies of the moving parts, we define spatio-temporal
relationships between pairwise components. The resulting descriptive
middle-level components and pairwise-components thereby catch the essential motion characteristics of human activities. They also give a very
compact representation of the video.
|
|
Visual features and data representation
[+/-]
Graph commute times for image representation
 |
|
We argue that the spatial layout needs to be taken into account in order to represent a
image content. This can be achieved by considering spectral properties of the feature graph constructed on
extracted interest points. To be more precise, we first construct the image feature graph of the image, then
the commute time matrix of the collapsed feature graph constitutes our image representation. An appropriate
machine learning algorithm is then used to perform image classification tasks.
|
[+/-]
Mixture distribution for weakly supervised classification in remote sensing images
 |
The general idea of the shape descriptor we introduce is to characterise the spatial
distribution of keypoints within a given region by computing the Fourier transform of the function of their
occurrences for a given radius. The descriptor is then described by the module of the first n coefficients of the
FT, hence making it invariant to rotation transformation. We explain it as a transformation of a generalised
correlogram taken at a keypoint.
|
Remote Sensing Applications : registration, segmentation and change analysis
[+/-]
Texture analysis for change detection
We define a measure of the observed change based on the distribution of the coefficients
issued from a wavelet transform, taking care to be rotation invariant.
The dissimilarities are
obtained through the Kullback-Liebler distance and a change
features vector is defined from all the distances between the
bands of the wavelet decomposition. This measurement is
able to classify the nature of the change between two images.
See Project page - Related article : [ ICIP 09 ] (COSTEL as PI)
|
|
[+/-]
Robust image matching and change detection from SVM classifier
|
The core ideas in this work are : i) to propose an image
change detection scheme based on
learning robust features computed at points of interest;
ii) to introduce a point matching algorithm that reduces
false-positives detection due to
high buildings projection effects.
|
|
Key words
Scene understanding, Image analysis, Computer Vision, Data Assimilation,
Statistical Learning, Earth observation, Applications.
