Robust Depth Map Estimation Using Motion Vector Representations – This study proposes a new technique for 3D reconstruction from partial deformation measurements in low- and high-resolution datasets. This is accomplished by constructing the partial measurements for each point in a deformable space, based on a mapping scheme of a data-rich optical flow. This mapping scheme is also exploited by extracting the high-resolution reconstruction from data. To tackle the problem of high-resolution deformation measurements, the proposed technique is first applied to a large dataset of deformable signals, and then combines the reconstructed partial measurements to improve the reconstruction performance. Experiments on simulated and real deformation measurements indicate that the proposed approach achieves comparable results compared to state-of-the-art methods.

Recent works show that deep neural network (DNN) models perform very well when they are trained with a large number of labeled samples. Most DNNs learn the classification model for each instance only and ignore the training data for classification. In this work we develop a probabilistic approach for training deep networks in such a way that the data are not being actively sampled. Our approach is based on combining the notion of model training and the notion of data representation by explicitly modeling the prior distribution over the data for the task of inferring the class of objects. As the model is learned with the distribution of the data in mind, the model is able to predict the model to be labeled, and to use the prediction of the model to infer the class of objects. We show that by using the distribution, the model can be trained to use the model to classify the objects with the most informative labels. Our proposed method is effective, general, and runs well on various high-scoring models of several real datasets.

PanoqueCa: Popper and Context for Semantic Parsing of Large Categorical Datasets

# Robust Depth Map Estimation Using Motion Vector Representations

Deep Learning Semantic Part Segmentation

A Deep Learning Approach for Image Retrieval: Estimating the Number of Units Segments are UnavailableRecent works show that deep neural network (DNN) models perform very well when they are trained with a large number of labeled samples. Most DNNs learn the classification model for each instance only and ignore the training data for classification. In this work we develop a probabilistic approach for training deep networks in such a way that the data are not being actively sampled. Our approach is based on combining the notion of model training and the notion of data representation by explicitly modeling the prior distribution over the data for the task of inferring the class of objects. As the model is learned with the distribution of the data in mind, the model is able to predict the model to be labeled, and to use the prediction of the model to infer the class of objects. We show that by using the distribution, the model can be trained to use the model to classify the objects with the most informative labels. Our proposed method is effective, general, and runs well on various high-scoring models of several real datasets.