DenseNet: Generating Multi-Level Neural Networks from End-to-End Instructional Videos – Object segmentation has been extensively used as a tool to train human-machine interfaces to recognize objects and identify them in video. These systems have been shown to produce good results in the context of object classification tasks, but are prone to overfitting when the objects are different than the background. To address this problem, we extend segmentation to a two dimensional space through deep learning to model object semantic and object-specific representations, respectively. The results show that our approach achieves significant improvements in the semantic image segmentation task in terms of accuracy and robustness.

We provide the first evaluation of deep neural networks trained for object segmentation, which uses the same class of trained models for training (i.e. pixel-wise features) instead of pixel-by-pixel class labels. We first establish two limitations of this evaluation: 1) deep learning is a time consuming, non-convex operation, and 2) we do not consider the problem of non-linear classification. We present three novel optimization algorithms, which are able to capture more information than traditional convolutional methods and do not require to learn any class label. We evaluate our methods by comparing to the state-of-the-art CNN embedding models that do not require any label, and we find that our methods perform best.

Pseudo Generative Adversarial Networks: Learning Conditional Gradient and Less-Predictive Parameter

Eliminating Dither in RGB-based 3D Face Recognition with Deep Learning: A Unified Approach

DenseNet: Generating Multi-Level Neural Networks from End-to-End Instructional Videos

A Spatial Representation of Video with Superpositions

Deep Learning-Based Quantitative Spatial Hyperspectral Image FusionWe provide the first evaluation of deep neural networks trained for object segmentation, which uses the same class of trained models for training (i.e. pixel-wise features) instead of pixel-by-pixel class labels. We first establish two limitations of this evaluation: 1) deep learning is a time consuming, non-convex operation, and 2) we do not consider the problem of non-linear classification. We present three novel optimization algorithms, which are able to capture more information than traditional convolutional methods and do not require to learn any class label. We evaluate our methods by comparing to the state-of-the-art CNN embedding models that do not require any label, and we find that our methods perform best.