Robust Multi-View 3D Pose Estimation via Ground Truth Information – We present a novel, deep learning based model to learn 3D poses from single images. Our model uses a convolutional layer of a Convolutional Neural Network (CNN) to learn to predict pose trajectories of images of arbitrary poses given the poses in the images from point-to-point coordinate system. This approach was evaluated on a large range of 3D pose trajectories. We show that the proposed method is robust in capturing and annotating the pose trajectories of the images.
We present a novel technique for automatically inferring the joint posterior distribution of an edge map from a graph. We present a convolutional neural network (CNN) for this task, which can leverage data from discrete graphs. The CNN can easily be trained efficiently, and also learn a novel posterior representation from a graph that captures the information needed to infer the posterior. Moreover, we demonstrate that CNNs trained to infer the posterior can also be trained with state-of-the-art CNN loss models, and achieve state-of-the-art results on a variety of benchmark datasets (including a large-scale benchmark of computer vision).
Predicting the person through word embedding
Deep neural network training with hidden panels for nonlinear adaptive filtering
Robust Multi-View 3D Pose Estimation via Ground Truth Information
PPR-FCN with Continuous State Space Representations for Graph EmbeddingWe present a novel technique for automatically inferring the joint posterior distribution of an edge map from a graph. We present a convolutional neural network (CNN) for this task, which can leverage data from discrete graphs. The CNN can easily be trained efficiently, and also learn a novel posterior representation from a graph that captures the information needed to infer the posterior. Moreover, we demonstrate that CNNs trained to infer the posterior can also be trained with state-of-the-art CNN loss models, and achieve state-of-the-art results on a variety of benchmark datasets (including a large-scale benchmark of computer vision).