Open Access. Powered by Scholars. Published by Universities.®
Articles 1 - 8 of 8
Full-Text Articles in Entire DC Network
Infomixup : An Intuitive And Information-Driven Approach To Robust Generalization, Andrew H. Meyer
Infomixup : An Intuitive And Information-Driven Approach To Robust Generalization, Andrew H. Meyer
Theses
The discovery of Adversarial Examples — data points which are easily recognized by humans, but which fool artificial classifiers with ease, is relatively new in the world of machine learning. Corruptions imperceptible to the human eye are often sufficient to fool state of the art classifiers. The resolution of this problem has been the subject of a great deal of research in recent years as the prevalence of Deep Neural Networks grows in everyday systems. To this end, we propose InfoMixup , a novel method to improve the robustness of Deep Neural Networks without significantly affecting performance on clean samples. …
Learning Representations In The Hyperspectral Domain In Aerial Imagery, Aneesh Rangnekar
Learning Representations In The Hyperspectral Domain In Aerial Imagery, Aneesh Rangnekar
Theses
We establish two new datasets with baselines and network architectures for the task of hyperspectral image analysis. The first dataset, AeroRIT, is a moving camera static scene captured from a flight and contains per pixel labeling across five categories for the task of semantic segmentation. The second dataset, RooftopHSI, helps design and interpret learnt features on hyperspectral object detection on scenes captured from an university rooftop. This dataset accounts for static camera, moving scene hyperspectral imagery. We further broaden the scope of our understanding of neural networks with the development of two novel algorithms - S4AL and S4AL+. We develop …
Deeprm: Deep Recurrent Matching For 6d Pose Refinement, Alexander Avery
Deeprm: Deep Recurrent Matching For 6d Pose Refinement, Alexander Avery
Theses
Precise 6D pose estimation of rigid objects from RGB images is a critical but challenging task in robotics and augmented reality. To address this problem, we propose DeepRM, a novel recurrent network architecture for 6D pose refinement. DeepRM leverages initial coarse pose estimates to render synthetic images of target objects. The rendered images are then matched with the observed images to predict a rigid transform for updating the previous pose estimate. This process is repeated to incrementally refine the estimate at each iteration. LSTM units are used to propagate information through each refinement step, significantly improving overall performance. In contrast …
A Depth-Based Computer Vision Approach To Unmanned Aircraft System Landing With Optimal Positioning, Nicholas Quattrociocchi
A Depth-Based Computer Vision Approach To Unmanned Aircraft System Landing With Optimal Positioning, Nicholas Quattrociocchi
Theses
High traffic congestion in cities can lead to difficulties in delivering appropriate aid to people in need of emergency services. Developing an autonomous aerial medical evacuation system with the required size to facilitate the need can allow for the mitigation of the constraint. The aerial system must be capable of vertical takeoff and landing to reach highly conjected areas and areas where traditional aircraft cannot access. In general, the most challenging limitation within any proposed solution is the landing sequence. There have been several techniques developed over the years to land aircraft autonomously; however, very little attention has been scoped …
Learning Multi-Step Robotic Manipulation Tasks Through Visual Planning, Sulabh Kumra
Learning Multi-Step Robotic Manipulation Tasks Through Visual Planning, Sulabh Kumra
Theses
Multi-step manipulation tasks in unstructured environments are extremely challenging for a robot to learn. Such tasks interlace high-level reasoning that consists of the expected states that can be attained to achieve an overall task and low-level reasoning that decides what actions will yield these states. A model-free deep reinforcement learning method is proposed to learn multi-step manipulation tasks. This work introduces a novel Generative Residual Convolutional Neural Network (GR-ConvNet) model that can generate robust antipodal grasps from n-channel image input at real-time speeds (20ms). The proposed model architecture achieved a state-of-the-art accuracy on three standard grasping datasets. The adaptability of …
Deep Feature Learning And Adaptation For Computer Vision, Abu Md Niamul Taufique
Deep Feature Learning And Adaptation For Computer Vision, Abu Md Niamul Taufique
Theses
We are living in times when a revolution of deep learning is taking place. In general, deep learning models have a backbone that extracts features from the input data followed by task-specific layers, e.g. for classification. This dissertation proposes various deep feature extraction and adaptation methods to improve task-specific learning, such as visual re-identification, tracking, and domain adaptation. The vehicle re-identification (VRID) task requires identifying a given vehicle among a set of vehicles under variations in viewpoint, illumination, partial occlusion, and background clutter. We propose a novel local graph aggregation module for feature extraction to improve VRID performance. We also …
Towards Efficient Lifelong Machine Learning In Deep Neural Networks, Tyler L. Hayes
Towards Efficient Lifelong Machine Learning In Deep Neural Networks, Tyler L. Hayes
Theses
Humans continually learn and adapt to new knowledge and environments throughout their lifetimes. Rarely does learning new information cause humans to catastrophically forget previous knowledge. While deep neural networks (DNNs) now rival human performance on several supervised machine perception tasks, when updated on changing data distributions, they catastrophically forget previous knowledge. Enabling DNNs to learn new information over time opens the door for new applications such as self-driving cars that adapt to seasonal changes or smartphones that adapt to changing user preferences. In this dissertation, we propose new methods and experimental paradigms for efficiently training continual DNNs without forgetting. We …
Multi-Scale Architectures For Human Pose Estimation, Bruno Artacho
Multi-Scale Architectures For Human Pose Estimation, Bruno Artacho
Theses
In this dissertation we present multiple state-of-the-art deep learning methods for computer vision tasks using multi-scale approaches for two main tasks: pose estimation and semantic segmentation. For pose estimation, we introduce a complete framework expanding the fields-of-view of the network through a multi-scale approach, resulting in a significant increasing the effectiveness of conventional backbone architectures, for several pose estimation tasks without requiring a larger network or postprocessing. Our multi-scale pose estimation framework contributes to research on methods for single-person pose estimation in both 2D and 3D scenarios, pose estimation in videos, and the estimation of multiple people’s pose in a …