In recent years, the rapid development of deep learning approaches has paved the way to explore the underlying factors that explain the data. In particular, several methods have been proposed to learn to identify and disentangle these underlying explanatory factors in order to improve the learning process and model generalization. However, extracting this representation with little or no supervision remains a key challenge in machine learning. In this paper, we provide a theoretical outlook on recent advances in the field of unsupervised representation learning with a focus on auto-encoding-based approaches and on the most well-known supervised disentanglement metrics. We cover the current state-of-the-art methods for learning disentangled representation in an unsupervised manner while pointing out the connection between each method and its added value on disentanglement. Further, we discuss how to quantify disentanglement and present an in-depth analysis of associated metrics. We conclude by carrying out a comparative evaluation of these metrics according to three criteria, (i) modularity, (ii) compactness and (iii) informativeness. Finally, we show that only the Mutual Information Gap score (MIG) meets all three criteria.

In order to improve behavioral analysis systems in urban environments, this paper proposes, using data extracted from video surveillance cameras, a tracking method through two approaches. The first approach consists in comparing the position of people between two images of a video and to perform tracking by proximity. The second method using Kalman filters is based on the anticipation of the position of an individual in the upcoming image. The use of this method proves to be more efficient as it allows continuing a detection even when people cross each other or when they pass behind obstacles. The use of Kalman filters in this domain provides a new approach to obtain reliable tracking and information on speed and trajectory variations. The proposed method is innovative in the way the tracking is performed and the results are exploited. Experiments were conducted in a real situation and showed that the use of some elements of the first method could be reused to integrate a notion of distance in the method based on the Kalman filter and thus improve the latter both in tracking and in detecting of abnormal behavior. This article deals with the functioning of the two methods as well as the results obtained with the same scenarios. The experimentation concludes through concrete results that the Kalman filter method is more efficient than the proximity method alone. A sample result is available online for two of the seven videos used in this article (accessed on 19 July 2021).

The richness of hyperspectral imaging (HSI)-collected signals makes possible quantitative inference by the inverse problem-solving of the chemical and biophysical parameters of the imaged object. In this article, we first propose a classification of the large variety of literature on invertible physics-based hyperspectral signature models by analyzing their founding hypotheses and methodologies. All the models found in the literature are based on the radiative transfer theory (RTT), but they are divided into three main branches. In the first one, the models rely on the Beer-Lambert law. In the second and the third ones, the models are based on Lommel’s radiative transfer equation (RTE), which they simplify into a diffusion approximation equation or into a single- and multiple-scattering approximation equation, respectively. Thereafter, we present the most recent models available for each branch for applications in geoscience and remote sensing domains: MARMIT-1 and MARMIT-2 for soil, the Microphytobenthos Optical Model (MPBOM) for algae and bacteria biofilm, PROSPECT for plant leaves, Farrell for fruits and vegetables, and Hapke for solar system objects. On the one hand, this article provides an overview of the different works in the literature. On the other hand, it proposes a classification tree of the different models, allowing one to know the similarities and differences of the different models. Ultimately, the aim is to assist researchers in selecting an appropriate model based on their specific targeted application.