Multi-label Visual Place Matching – A major challenge in the area of Convolutional Neural Networks (CNN) is the lack of explicit representation of multiple target regions. In this work, we present a novel method which enables the learning of multiple target regions without supervision (i.e., labeling) at each instant. The method is based on a novel combination of an external feature learning algorithm and a novel network architecture, which is based on a local and long-term memory network. Experimental evaluation on two different datasets, MNIST and MS-DB, reveals that our method outperforms the state-of-the-art CNN architectures on the MNIST dataset to the same extent as or better than the state of the art models.
This paper presents a new unsupervised feature learning algorithm for high-dimensional structured labels, such as those generated by large image sensors. By using a single feature model, the discriminator of each label can be predicted with a maximum likelihood estimate as well as a maximum likelihood of features that correspond to the data points. The problem is solved by a novel deep-learning based algorithm which combines the effectiveness of a feature classifier and a single label classifier. Experiments show that the algorithm compares favorably with state-of-the-art deep learning algorithms.
A Novel Face Alignment Based on Local Contrast and Local Hue
A statistical approach to statistical methods with application to statistical inference
Multi-label Visual Place Matching
Efficient Semidefinite Parallel Stochastic Convolutions
Efficient Hierarchical Clustering via Deep Feature FusionThis paper presents a new unsupervised feature learning algorithm for high-dimensional structured labels, such as those generated by large image sensors. By using a single feature model, the discriminator of each label can be predicted with a maximum likelihood estimate as well as a maximum likelihood of features that correspond to the data points. The problem is solved by a novel deep-learning based algorithm which combines the effectiveness of a feature classifier and a single label classifier. Experiments show that the algorithm compares favorably with state-of-the-art deep learning algorithms.