The A3SI team of the LIGM (Joint Research Unit of the University of Paris East) has been collaborating with several Brazilian universities for more than 20 years. On the occasion of the visit to France of Arnaldo de Albuquerque Araújo (post-doc in the team in 1994), a Franco-Brazilian workshop of image analysis is organized Thursday, December 8 from 10:00 to 12:00, amphi 110 (ESIEE PARIS).
Six doctoral students will present their work in a Franco-Brazilian research context:
Hierarchies of segmentations have been extensively studied in the past few decades and their evaluation on user annotated image datasets shows their potential to further researches. Hierarchies of segmentations can be described by saliency maps, which weights the borders of a segmentation according to the levels they belong to. In this work, we combine hierarchical watershed image segmentations through their saliency maps. The combinations are split in two classes: global combinations, where all borders in each saliency map are uniformly handled, and concatenations, which simulate mergings of hierarchies at distinct levels. Our results show that some combined hierarchies perform significantly better than individual hierarchies.
In recent years, deep architectures have been used for transfer learning with state-of-the-art performance in many datasets. The properties of their features remain, however, largely unstudied under the transfer perspective: are they resilient? redundant? compressible? To answer these questions, we introduce simple perturbations to image descriptions extracted from a deep convolutional neural network, measuring how it affects the performance of a classifier trained on them, and show that they are more robust to these disturbances when compared to classical approaches, achieving remarkable scores even when heavily compressed.
This talk presents the systems developed by PUC Minas and IRISA for the person discovery task at MediaEval 2016. The task of person discovery in broadcast TV consists in identifying and naming persons of a video corpus which both speak and are visible at the same time, in an unsupervised way. The approach presented, used a graph-based representation along with tag-propagation techniques to associate overlays co-occurring with some speaking faces to other visually or audio-visually similar speaking faces. In this work, two tag-propagation methods were considered, one based on a random walk strategy, and the other on Kruskal’s algorithm. By using this approach, the team formed by PUC Minas and IRISA ended in second place, and received a distinctive mention at the MediaEval 2016 workshop.
Vision-based human action recognition is the process of labeling image sequences (e.g., video) with action labels. The image sequences provide information of color (RGB), and depending on the sensor used, it can also provide depth information. Human action recognition is essential in computer vision to automatically understand the action or actions contained in a scene. It is often tackled in three steps: features extraction, feature-based image modeling and classification. The main goal of this PhD thesis is to investigate and to assess the possibilities of action recognition performed with the help of mathematical morphology. We will be focused on providing efficient mathematical morphology feature extraction algorithms suitable for action recognition.
Many problems in image processing can be modelled by differential equations and it is not rare to find similarites among models for different problems. For example, image filtering and restauration, tomographic reconstruction, classification and segmentation of image can be solved in a common framework that counts with a large class of efficient algorithms. Among these, discrete calculus presents some advantages that were not yet explored for this class of problems, as fast convergence and suitable metrics for the subjacent problem. The goal of this thesis is to generalize the use of discrete calculus to a larger class of variational problems in image processing. The expressiveness of discrete calculus will be used to propose new regulators, mainly geometrical, using the curvature for example, and also new dedicated optimization algorithms.
Scene annotation consists in delineating and providing semantic labels to objects of interest that appear in a scene. The scene can be an image or a sequence of images (video). This problem is essential in computer vision to automatically understand the content of a scene. It is often tackled in three steps: segmentation, detection and recognition. The main goal of this PhD thesis is to investigate and to determine the possibilities of video annotation performed with the help of mathematical morphology. More specifically, we aim to work on incremental mathematical morphology algorithms which are suitable for efficient video segmentation.