Stacked Neural Networks for Semantic Segmentation of Hyperspectral Images – This paper addresses the challenge of estimating the spatial information of a semantic visual object using object motion data, such as images with multiple objects in a single motion-dependent frame. We propose a novel method called spatially noisy object detection (SPD) for semantic visual object detection, and demonstrate improved results by integrating object motion with motion data from a mobile robot. The SPD method exploits a convolutional neural network to construct a temporally noisy semantic image of a semantic object under a given object motion background scene. The proposed method is evaluated using images from a human-computer interaction task of object-tracking and recognition, and two videos captured directly from a robot using the robot’s smartphone. In all experiments, SPD achieved high object detection accuracy while maintaining a frame-accurate semantic visual representation. In the most difficult cases, it achieved superior results over the conventional method. In all tests, SPD outperformed the existing state-of-the-art methods, using a frame-accurate representation for object motion.
We investigate supervised deep learning for visual tracking. We propose a technique that extracts a representation of the sensor-dependent motion of the object and a neural network that uses a convolutional neural network to predict the appearance and orientation of the object accordingly. This representation can be used by using a convolutional neural network based on object-view-label pairs. We design and test a deep tracking system to accurately track a pair of objects. Through experimental evaluation, we demonstrate the effectiveness of our approach and demonstrate the effectiveness of our system on various real-world datasets.
Uncertainty Decomposition in Belief Propagation
Facial-torture reconstruction with deep convolutional autoencoders
Stacked Neural Networks for Semantic Segmentation of Hyperspectral Images
A General Algorithm for Grouping Visual Features into Semantic Spaces
Super-Dense: Robust Deep Convolutional Neural Network Embedding via Self-Adaptive RegularizationWe investigate supervised deep learning for visual tracking. We propose a technique that extracts a representation of the sensor-dependent motion of the object and a neural network that uses a convolutional neural network to predict the appearance and orientation of the object accordingly. This representation can be used by using a convolutional neural network based on object-view-label pairs. We design and test a deep tracking system to accurately track a pair of objects. Through experimental evaluation, we demonstrate the effectiveness of our approach and demonstrate the effectiveness of our system on various real-world datasets.