---

3D surface compression is a selected method for reduction of required memory and effective triangular mesh data transfer in networks with low bandwidth. In this paper, an improved technique for compression of triangular surfaces is introduced. It is based on centroidal Voronoi tessellation with density function. By choosing appropriate density function with curvature feature and Lloyd relaxation, vertex density in compressed mesh tends toward details in rough surfaces and prevents vertex redundancy and non-necessary vertex concentration in smoother surface areas. In post-processing step, non-linear Nelder-Mead optimization is applied for better vertex localization and reducing compression error in the simplified mesh. Our proposed method is compared with classic and modern techniques in recent studies. Implementation results show improvements in compression and accuracy of our method compared with available techniques.

---

3D human pose estimation is one of the most significant tasks in computer vision with wide range of applications. The works for estimating human pose initialized from 2D skeletal estimation from multiple data and has proceeded toward 3D skeletal estimation from minimum input information. In this paper, 3D human pose estimation from a single RGB image is investigated. The proposed work is considered as the ones which firstly estimate 2D pose and then lift the estimated 2D configuration to 3D space. Since most of the errors in this attitude are originated by inaccurate 2D pose inference, we have proposed a method for predicting more accurate 2D poses to obtain 3D poses with less errors. The proposed approach for estimating 2D pose has leveraged deep learning along with the information of the edge map. In other words, we have made use of edge features, which are hand-designed features, in order to guide the deep neural network in training and in learning the features in accordance with the defined objective. Experimental results have demonstrated less errors in 2D and consequently 3D pose estimation in Human3.6M and HumanEva-I benchmarks.

---

Human facial generation of example image is used as a requirement for biometric applications for the purpose of identifying individuals. In this paper, face generation consists of three main steps. In the first step, detection of significant lines and edges of the example image are carried out using nonlinear grayscale morphology. Then, hair areas are identified from the face of sample. The final step combines images from previous steps. Similarity and matching between synthesized face sketch and artistic sketch are compared with two methods of extracting features, Principle Component Analysis and Linear Discriminant Analysis, and time of the process is calculated. The experiments on the pair of CUHK database images show that the proposed method compared with state of the art methods such as: Eigen transformation, LLE, and MRF, has no computational complexity and creates a person's face with good quality and much less time. Matching of synthesized face sketch of the proposed method is achieved with a maximum value of 90% when Linear Discriminant Analysis is used to extract feature. The proposed method is also resistant to background effects and brightness of example images.

---

Sign language recognition systems help deaf people to access various media. In this paper, the Leap Motion Controller (LMC) and the image of the hand are exploited for sign language recognition. The LMC provides 3D position of the hand joints. The first set of features are extracted from the data provided by the LMC. When the hand is not located in vertical view of the LMC or when the hand posed like a fist, the precise position of the hand joints is not recognizable. The second feature extracted from hand image helps most hand gestures be recognized precisely. The second feature includes histogram of oriented gradients and the distance of the hand contour form the center of the hand. Also, a dataset composed of variant American sign language gestures is created which includes 64000 samples. In recognition stage, random forest is applied which is a good option for large datasets. The experimental results show that the proposed method performs better than similar methods.
 

---

Nowadays, various kind of smart phones and 3D software are produced, which require large memory space. However, large number of vertices and faces in 3D models not only decrease the speed of sending and receiving of data but also can make problem in systems with low memory space. In this paper, an anisotropic re-meshing of 3D models is proposed. In this algorithm, the Nyquist theorem is employed for sampling from each selected segment of the mesh, locally. Then, the re-triangulation algorithm is applied to the selecte samples to construct the simplified mesh. In order to construct a high quality mesh from the remeshed model, a non linear subdivision is employed. The achieved results show that the algorithm can reduce the number of vertices and faces beside preserving details of model. The proposed method is also compared to the state-of-the-art algorithms are used in simplification studies, the outcomes illustrate the ability of the proposed method in producing high quality models.
 

---

Face plays an important role in visual communication. By looking at the face, it can be automatically extracted many non-verbal messages, such as identity, intention, and emotion. In computer vision, localization of the key points of the face is usually a key step for automatic extraction of face information, and many facial analysis techniques are built on the precise recognition of these embossed. Facial landmark detection and alignment in images with occlusion is a very important and challenging task in many visual and image processing tasks. In this paper, a comprehensive method for initialization and alignment of facial landmark through training of local binary features (LBP) and histogram orientated gradient (HOG) and a facial landmark detection method using robust cascade pose regression, which are specified as pixel difference features of landmarks, is introduced. At first, by analyzing the correlation of the local binary pattern histogram (LBP) and then by using histogram orientated gradient, the features of the training images are obtained. For the test image using these features the instructional images are estimated as optimal guide points. In the test stage, according to initialization of the image, the selection of the appropriate feature for the image is used to speed up the process, which means the number of steps to be chosen for each image is better. A strong cascade mode regression is then used to adjust the face, and a local principle is applied to learn the features of the guide points. The local principle helps to learn a set of highly distinctive binary features for the face guide points independently; these local binary features are used to jointly learn the cascade mode regression for the final output. The results show that the initialization used in this work has increased the accuracy of the estimation in the cascade state regression and has obtained better results than the random initialization.