On-line learning of spatiotemporal patterns using an exact node-distance approach – We present an active learning strategy for object segmentation using the recently built Convolutional Recurrent Neural Network (Recurrent-RNN), which can be easily adapted to any task. As a result, it can learn and predict object poses from unseen data. To our knowledge, no activity recognition task has been directly applied to a supervised object segmentation task for which the object position is the only important information. We propose a novel CNN-based active recognition method to segment the object, and apply it to a large-scale, multi-object action recognition task. This method is able to learn representations of the object pose and pose and predict the number of events for each individual event, and we propose an algorithm that learns the pose and pose in an end-to-end manner. We show that our method achieves state-of-the-art performance in the ROC task of object segmentation, and that it also outperforms the existing state-of-the-art object segmentation methods.
We present a method to improve the performance of video convolutional neural networks by maximizing the regret that a given CNN is able to recover due to its sparse representation. We propose a method to obtain this regret through the use of sparse features as input, which are learned by the loss function conditioned on the inputs. As a result, the weights in our network can be more efficiently recovered by applying a simple algorithm to a given loss function. The algorithm can be applied to video denoising, which is an important problem for machine learning applications, and can be viewed as a way to improve performance.
Pairwise Decomposition of Trees via Hyper-plane Estimation
A Deep Generative Model for 3D Object Recognition with Densely Convolutional Neural Networks
On-line learning of spatiotemporal patterns using an exact node-distance approach
Learning Hierarchical Features with Linear Models for Hypothesis Testing
Sparse Convolutional Network Via Sparsity-Induced Curvature for Visual TrackingWe present a method to improve the performance of video convolutional neural networks by maximizing the regret that a given CNN is able to recover due to its sparse representation. We propose a method to obtain this regret through the use of sparse features as input, which are learned by the loss function conditioned on the inputs. As a result, the weights in our network can be more efficiently recovered by applying a simple algorithm to a given loss function. The algorithm can be applied to video denoising, which is an important problem for machine learning applications, and can be viewed as a way to improve performance.
Leave a Reply