Computing Science - Theses, Dissertations, and other Required Graduate Degree Essays

Receive updates for this collection

UAV object-based semi-autonomous and autonomous navigation

Date created: 
2021-07-19
Abstract: 

In this thesis, we have developed a semi-autonomous behavior that allows us to control a drone with less effort. We have also presented a technique that enables autonomous repeating of a previously traversed route using the visual navigation system. Our first application demonstrates an experiment with driving a drone using a vision-based control (visual servoing) method, particularly by tracking selected targets in an image view. In the second application, a drone equipped with a monocular camera has been derived manually on a path. Invariant semantic features (i.e., objects) have been extracted using an object detection neural network, YOLO. Using these features, we show that the drone can repeat the traversed route autonomously independently from the lighting condition and even appearance changes.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Richard Vaughan
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.

On real-time data fusion in edge computing

Author: 
Date created: 
2021-08-27
Abstract: 

Recent years have witnessed a drastic increase in the scale of data generated by sensors and smart devices across the city. While the data scale is increasing, it is more and more important to process the data in a real-time manner. Numerous new applications are to be enabled by low latency real-time processing. The inevitable transmission delay in cloud computing brings the new computing paradigm "edge computing'', which aims to locate computing resources near the end users. Meanwhile, the inevitable weak computing power of edge servers brings challenges to information retrieval quality. In this thesis, we explored multimodality/sensor fusion solutions to enable new applications and to optimize the latency-accuracy trade-off in resource-limited edge scenarios. We first presented a synchronous multimodality stream analytics framework with a typical use case: profanity filtering in real-time video conference. We implemented and evaluated our prototype by real-world scenario test cases. Our system achieves good profanity filtering rate (89%) while maintaining the synchronicity of the video stream and not affecting the overall latency (400 ms), which indicates the potential of multimodality stream processing for new applications in resource-limited scenarios. We then presented a high-accuracy low-latency road information collection system based on object-level fusion. By a multi-path resistant design, our prototype system outperforms not only a visual-only solution but also a state-of-the-art camera-radar sensor fusion solution. Extensive real-world evaluation shows that our system can reduce 25% of the localization error and increase 45\% of the recognition rate comparing with a state-of-the-art method, which confirms the great potential of our method in achieving high-accuracy low-latency object information retrieval.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Jiangchuan Liu
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.

Transforming neural machine translation into simultaneous text and speech translation

Author: 
Date created: 
2021-07-29
Abstract: 

Simultaneous neural Machine Translation (SiMT) aims to maintain translation quality while minimizing the delay between reading the input and incrementally producing the output. The eventual goal of SiMT is to match the performance of highly skilled human interpreters who can simultaneously listen to a speaker in a source language and produce a translation in the target language with minimal delay. In this thesis, we explore attempts at building reliable simultaneous translation systems that can produce fluent translations with minimal latency. We present two distinct methods for finding an optimal policy that tells us if current input is enough for generating accurate translations, or we need to wait for more information. Our first method employs a prediction mechanism to inform the model about incoming input stream. We show as the length of sentences grows, predicting future time steps become essential due to more complex re-orderings that can happen more often in long sentence pairs. Our second method introduces a new algorithmic approach for finding optimal policy as a reference in a supervised learning model. The resulting system translates more accurately with less delay. Our third project focuses on improving the performance of an end-to-end speech translation system, which many simultaneous speech systems rely on. We propose a new loss function that allows us to use available huge datasets for Machine Translation task in order to improve the performance of speech translation system.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Anoop Sarkar
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) Ph.D.

Energy analysis and optimization for distributed data centers under electrical load shedding

Author: 
Date created: 
2021-08-12
Abstract: 

The number and scales of data centers have significantly increased in the current digital world. The distributed data centers are standing out as a promising solution due to the development of modern applications which need a massive amount of computation resources and strict response requirements. Their reliability and availability heavily depend on the electrical power supply. Most of the data centers are equipped with battery groups as backup power in case of electrical load shedding or power outage due to severe weather or human-driven factors. The limited numbers and degradation of batteries, however, can hardly support the servers to finish all the jobs on time. In this thesis, we divide all the workload in data centers into web jobs and batch jobs. We develop a battery allocation and workload migration framework to guarantee web jobs are never interrupted and try to minimize the waiting time of batch jobs simultaneously. Our extensive evaluations show that our battery allocation and workload migration results can guarantee all the web jobs uninterrupted and minimize the average waiting time of batch jobs within a limited overall cost compared to the current practical allocation.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Jiangchuan Liu
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.

An efficient oracle for counting shortest paths in planar graphs

Author: 
Date created: 
2021-08-09
Abstract: 

An O(\sqrt {n}) query time and O(n^{1.5}) size oracle for counting shortest paths is proposed. Given a pair of vertices u and v in a planar graph G of n vertices, the oracle answers the number of shortest paths from u to v in O(\sqrt {n}) time and whether there is a unique shortest path from u to v in O(\log n) time. Bezáková and Searns [ISAAC 2018] give an O(\sqrt {n}) query time and O(n^{1.5}) size oracle for counting shortest paths in planar graphs. Applying this oracle directly, it takes O(\sqrt {n}) time to answer whether there is a unique shortest path from u to v. A key component in our oracle is to apply Voronoi diagrams on planar graphs, which is a recent novel notion used in oracles for answering shortest distance, to speed up the query time. Computational studies show that our oracle is faster to answer queries than the oracle of Bezáková and Searns for large graphs. Our computational studies also confirm that Voronoi diagrams are efficient data structures for shortest distance oracles in practice.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Qianping Gu
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.

Conditional irrelevance in belief change

Author: 
Date created: 
2021-07-30
Abstract: 

This thesis presents an approach to incorporating qualitative assertions of conditional irrelevance into belief change, in order to address the limitations of existing work which considers only unconditional irrelevance. These assertions serve to enforce the requirement of minimal change to existing beliefs, while also suggesting a route to reducing the computational cost of belief change by excluding irrelevant beliefs from consideration. Our approach uses modified multivalued dependencies to represent domain-dependent conditional irrelevance assertions. We consider these assertions as capturing a property of the underlying domain, and consequently assume that a knowledge engineer has specified a collection of conditional irrelevance assertions to be taken into account during belief change. We introduce two related notions of what it means for a conditional irrelevance assertion to be taken into account by a belief revision or contraction operator: partial and full compliance. We also show that partially (and fully) compliant belief revision and contraction operators are interdefinable via the Levi and Harper identities. Further, we provide characterisations of partially and fully compliant belief revision operators in terms of semantic conditions on their associated faithful rankings. Using these characterisations, we show that partially and fully compliant belief revision operators exist. Finally, we compare our approach to existing work on unconditional irrelevance in belief change.

Document type: 
Thesis
File(s): 
Supervisor(s): 
James Delgrande
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.

Plan2Scene: Converting floorplans to 3D scenes

Date created: 
2021-07-27
Abstract: 

We address the Plan2Scene task: converting a floorplan and associated photos into a textured 3D mesh model of a residence. Our method 1) lifts a floorplan image to a 3D mesh, 2) synthesizes textures for observed surfaces based on input photos, and 3) generates textures for unobserved surfaces using a graph neural network architecture. We address the challenge of producing tileable textures for all architectural surfaces (floors, walls, and ceilings) from a sparse set of photos that only partially cover a house. To train and evaluate our system, we curate two texture datasets and extend a dataset of floorplans + photos from prior work with rectified surface crops and additional annotations. Our system produces realistic 3D models that outperform baseline approaches, as identified by a holistic user study and quantified by a suite of texture quality metrics. We release all our code, data, and trained models to the community.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Manolis Savva
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.

From estimation to control for robotic navigation: Probabilistic and optimal approaches

Author: 
Date created: 
2021-08-03
Abstract: 

Nowadays, mobile robots capable of autonomous navigation and interaction in unfamiliar and dynamic environments have received great attention among researchers. The robot must be able to precisely perceive its environment, make appropriate inference, plan its path, and travel around safely in order to achieve this goal. In robotics, maneuvering in a complex setting has been challenging. Several methods propose robust architectures in which the agent acts conservative in respect to uncertainty by considering worst case scenario, while others provide adaptive policies which try to adjust the actions given the concurrent knowledge. The usually suffers from guaranteed stability and efficiency in data. The novelty in this report is two folded: the first is to suggest a probabilistic framework for estimating environmental hazards and dynamic models. By improving MCMC, we propose an online method to obtain the model parameters distribution. The second novelty, is to propose an inference model and update framework for human navigational intent. We will discuss how one can apply these insights in a safe path planning problem by considering the environment's uncertainty in a probabilistic manner.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Mo Chen
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.

Articulated object reconstruction from interaction videos

Author: 
Date created: 
2021-08-10
Abstract: 

This thesis studies the problem of articulated object reconstruction from an input video. Our focus is on estimating the shape, pose, and part motion of an articulated object during human-object manipulation. The task is challenging as the object is dynamically changing and 3D reconstruction from 2D is inherently ambiguous. To enable research in this direction, we first create D3D-HOI: a dataset of monocular videos with ground truth annotations of 3D object shape, pose and part motion from human-object interaction videos. Our dataset consists of several common categories of articulated objects in diverse real-world scenes, observed from a variety of fixed camera view points. Each manipulated object (e.g., microwave) is represented using a 3D parametric model that best fits the captured data. We then annotate the size, pose, and part articulation values at every frame. A novel optimization-based method is proposed based on differentiable renderer and human-object interaction terms, which leverage the human pose for better inferring of the object spatial layout and dynamics. We evaluate this new approach on our dataset, demonstrating that human-object relations can significantly reduce the pose and motion errors on real-world articulated objects. Code and dataset are available at the following link (https://github.com/facebookresearch/d3d-hoi).

Document type: 
Thesis
File(s): 
Supervisor(s): 
Yasutaka Furukawa
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.

Scalable statistical-relational model discovery

Author: 
Date created: 
2021-08-04
Abstract: 

Many organisations store large amounts of data in relational databases and require efficient ways to extract useful information from them. Machine learning models learned from these databases enable intelligent queries to be answered. Typically these models require sufficient statistics in the form of frequency counts, which are efficiently captured by a contingency table (ct-table). Several techniques have been developed to generate ct-tables from a single table; however, in the case of multi-relational databases, unique challenges arise making these solutions inappropriate to use. In particular, the data is spread across multiple tables and must be joined to determine the correct frequency counts. In addition, counts for the non-existing relationships must be inferred as they are not explicitly stored in the database. This thesis presents a novel hybrid-counting (HYBRID) approach to computing ct-tables from relational databases that combines pre-counting (PRECOUNT) and post-counting (ONDEMAND) methods to provide a technique that is able to address the weaknesses in both methods.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Oliver Schulte
Department: 
Applied Sciences: School of Computing Science
Thesis type: 
(Thesis) M.Sc.