Literaturnachweis - Detailanzeige
| Autor/in | Yuan, Shuaihang |
|---|---|
| Titel | View-Based 3D Geometric Learning |
| Quelle | (2023), (183 Seiten)
PDF als Volltext Ph.D. Dissertation, New York University Tandon School of Engineering |
| Sprache | englisch |
| Dokumenttyp | gedruckt; online; Monographie |
| ISBN | 979-8-3684-5518-1 |
| Schlagwörter | Hochschulschrift; Dissertation; Computer Simulation; Geometry; Artificial Intelligence; Data Analysis; Visual Aids; Cues; Geometric Concepts; Recognition (Psychology); Learning Processes |
| Abstract | Recently, with the advancement in 2D imaging techniques and 3D visual sensors such as LiDAR, RGB-D cameras, etc. The use of 2D and 3D data is ubiquitous in various fields like autonomous driving, AR, and VR. Therefore, we are faced with an ever-increasing demand for approaches toward the automatic processing and analysis of data from multiple modalities. With the thriving of deep learning, various research works have been proposed to analyze data in a single domain. However, how to effectively process and analyze data by leveraging cross-domain information is less studied, which has become a challenging yet important research topic in the computer vision field. Existing works in the context of 3D geometric learning from 2D visual cues require various annotated 2D views to extract useful underlying geometric features. In this thesis, we focus on the problem of 3D geometric learning from 2D visual cues (i.e., RGB images, depth images, and sketches) by addressing the problem of how to effectively leverage geometric information covered in 2D views and how to alleviate the dependence of the annotated data. My methods can be categorized into four parts: 1) 3D shape recognition by learning from dynamic views. 2) 3D shape reconstruction by learning 2D and 3D distributions. 3) 3D shape detection by learning from few-shot training views. In the first part of the thesis, I propose to learn task-specific optimal 2D views from multi-view images for 3D shape recognition. In the multi-view based methods, it is critical to select the most appropriate viewpoints that convey discriminative regions of 3D shapes for different tasks. Existing methods often use a set of static views observed from pre-defined viewpoints without considering downstream tasks. In this method, we propose a new differentiable view learning scheme that enables task-specific dynamic viewpoint determination by considering the downstream shape recognition tasks. In addition to 3D shape recognition, I propose a method to reconstruct 3D shapes from 2D views. We observe that existing reconstruction methods fail to consider 1) the probabilistic distribution of the 2D images for given 3D objects and 2) the probabilistic distribution of 3D objects for given 2D images while learning to reconstruct 3D objects from single 2D images. This observation inspires us to propose a novel learning paradigm to fully utilize the generative representations to address these two critical factors. Then, I address the problem of few-shot indoor 3D object detection by proposing a meta-learning-based framework that only relies on a few labeled samples from novel classes for training. Given a query 3D point cloud and a few support samples, my method is trained over different 3D detection tasks to learn how to detect different object classes and dynamically adapt the 3D object detector to complete a specific detection task for the novel object by weighting the scene feature using a few support samples. [The dissertation citations contained here are published with the permission of ProQuest LLC. Further reproduction is prohibited without permission. Copies of dissertations may be obtained by Telephone (800) 1-800-521-0600. Web page: http://www.proquest.com/en-US/products/dissertations/individuals.shtml.] (As Provided). |
| Anmerkungen | ProQuest LLC. 789 East Eisenhower Parkway, P.O. Box 1346, Ann Arbor, MI 48106. Tel: 800-521-0600; Web site: http://www.proquest.com/en-US/products/dissertations/individuals.shtml |
| Erfasst von | ERIC (Education Resources Information Center), Washington, DC |
| Update | 2024/1/01 |
Literaturbeschaffung und Bestandsnachweise in Bibliotheken prüfen
Permalink als QR-Code
Inhalt auf sozialen Plattformen teilen (nur vorhanden, wenn Javascript eingeschaltet ist)
Teile diese Seite:
