Boosting and Deblurring with a Convolutional Neural Network – Feature extraction and classification are two important applications of machine learning in computer vision. In this work, we propose a novel deep convolutional neural network architecture called RNN-CNet to automatically train image classifiers. The RNN architecture is based on a CNN architecture, and is capable of handling the state-of-the-art convolutional neural networks. We demonstrate that the RNN-CNet is much more robust to the amount of labeled data than their CNN counterparts, with the advantage being that it can easily provide a compact representation of the class, which could be easily adapted for various applications. We also present a novel feature extraction technique to automatically predict the appearance of the objects that they occlude. The proposed approach is also evaluated on the task of object object pose estimation, and outperforms all other supervised CNN based methods on both benchmark and real-world datasets. We further demonstrate that the proposed feature extraction method outperforms all state-of-the-art CNN based model choices in three challenging datasets.

A key challenge of object detection and tracking in virtual environments is the need to identify the physical appearance of objects. Here, we present a novel architecture and dataset for the automatic classification of complex physical objects (e.g., faces and limbs). By leveraging the spatial-temporal structure between a virtual object and its physical appearance, the two tasks are unified into a multi-object class problem. With this framework, we further leverage temporal information in the appearance of a given object to improve object and object tracker performance. The performance of this architecture is evaluated on two real-world datasets, showing that the proposed architecture significantly improves tracking performance.

Stochastic gradient descent with two-sample tests

A Novel Feature Selection Method Using Backpropagation for Propositional Formula Matching

Boosting and Deblurring with a Convolutional Neural Network

A unified and globally consistent approach to interpretive scaling

The Eye of the Beholder: Learning Temporal Representation and Appearance for Action RecognitionA key challenge of object detection and tracking in virtual environments is the need to identify the physical appearance of objects. Here, we present a novel architecture and dataset for the automatic classification of complex physical objects (e.g., faces and limbs). By leveraging the spatial-temporal structure between a virtual object and its physical appearance, the two tasks are unified into a multi-object class problem. With this framework, we further leverage temporal information in the appearance of a given object to improve object and object tracker performance. The performance of this architecture is evaluated on two real-world datasets, showing that the proposed architecture significantly improves tracking performance.