This work is aimed at the detection of adult images that appear in Internet. Skin detection is of the paramount importance in the detection of adult images. We build a maximum entropy model for this task. This model, called the First Order Model in this paper, is subject to constraints on the color gradients of neighboring pixels. Parameter estimation as well as optimization cannot be tackled without approximations. With Bethe tree approximation, parameter estimation is eradicated and the Belief Propagation algorithm permits to obtain exact and fast solution for skin probabilities at pixel locations. We show by the Receiver Operating Characteristics (ROC) curves that our skin detection improves the performance over previous work [6] in the context of skin pixel detection rate and false positive rate. The output of skin detection is a grayscale skin map with the gray level indicating the belief of skin. We then calculate 9 simple features from this map which form a feature vector. Most of these features are based on fit ellipses, which are used to catch the characteristics of detected skin regions. Two fit ellipses are used for each skin map-the fit ellipse of all skin regions and the fit ellipse of the largest skin region. They are called respectively Global Fit Ellipse and Local Fit Ellipse in this paper. A multi-layer perceptron classifier is then trained for these features. Plenty of experimental results are presented, including photographs and a ROC curve calculated over a test set of 5, 084 photographs, which show stimulating performance for such simple features.

We consider a sequence of three models for skin detection built from a large collection of labelled images. Each model is a maximum entropy model with respect to constraints concerning marginal distributions. Our models are nested. The first model, called the baseline model is well known from practitioners. Pixels are considered as independent. Performance, measured by the ROC curve on the Compaq Database is impressive for such a simple model. However, single image examination reveals very irregular results. The second model is a Hidden Markov Model which includes constraints that force smoothness of the solution. The ROC curve obtained shows better performance than the baseline model. Finally, color gradient is included. Thanks to Bethe tree approximation, we obtain a simple analytical expression for the coefficients of the associated maximum entropy model. Performance, compared with previous model is once more improved.

In this paper, we tackle the problem of action recognition using body skeletons extracted from video sequences. Our approach lies in the continuity of recent works representing video frames by Gramian matrices that describe a trajectory on the Riemannian manifold of positive-semidefinite matrices of fixed rank. Compared to previous work, the manifold of fixed-rank positive-semidefinite matrices is endowed with a different metric, and we resort to different algorithms for the curve fitting and temporal alignment steps. We evaluated our approach on three publicly available datasets (UTKinect-Action3D, KTH-Action and UAV-Gesture). The results of the proposed approach are competitive with respect to state-of-the-art methods, while only involving body skeletons.

In this paper, we propose a method for 3D model indexing based on 2D views, named AVC (Adaptive Views Clustering). The goal of this method is to provide an optimal selection of 2D views from a 3D model, and a probabilistic Bayesian method for 3D model retrieval from these views. The characteristic views selection algorithm is based on an adaptive clustering algorithm and using statistical model distribution scores to select the optimal number of views. Starting from the fact that all views do not contain the same amount of information, we also introduce a novel Bayesian approach to improve the retrieval. We finally present our results and compare our method to some state of the art 3D retrieval descriptors on the Princeton 3D Shape Benchmark database.

No abstract