Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
An empirical assessment of real-time progressive stereo reconstruction
3D reconstruction from images, the problem of reconstructing depth from images, is one of the most well-studied problems within computer vision. In part because it is academically interesting, but also because of the significant growth in the use of 3D models. This growth can be attributed to the de...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
School of Architecture, Planning and Geomatics
2017
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613285104222208 |
|---|---|
| access_status_str | Open Access |
| author | Westaway, Matthew |
| author2 | Sithole, George |
| author_browse | Sithole, George Westaway, Matthew |
| author_facet | Sithole, George Westaway, Matthew |
| author_sort | Westaway, Matthew |
| collection | Thesis |
| description | 3D reconstruction from images, the problem of reconstructing depth from images, is one of the most well-studied problems within computer vision. In part because it is academically interesting, but also because of the significant growth in the use of 3D models. This growth can be attributed to the development of augmented reality, 3D printing and indoor mapping. Progressive stereo reconstruction is the sequential application of stereo reconstructions to reconstruct a scene. To achieve a reliable progressive stereo reconstruction a combination of best practice algorithms needs to be used. The purpose of this research is to determine the combinat ion of best practice algorithms that lead to the most accurate and efficient progressive stereo reconstruction i.e the best practice combination. In order to obtain a similarity reconstruction the in t rinsic parameters of the camera need to be known. If they are not known they are determined by capturing ten images of a checkerboard with a known calibration pattern from different angles and using the moving plane algori thm. Thereafter in order to perform a near real-time reconstruction frames are acquired and reconstructed simultaneously. For the first pair of frames keypoints are detected and matched using a best practice keypoint detection and matching algorithm. The motion of the camera between the frames is then determined by decomposing the essential matrix which is determined from the fundamental matrix, which is determined using a best practice ego-motion estimation algorithm. Finally the keypoints are reconstructed using a best practice reconstruction algorithm. For sequential frames each frame is paired with t he previous frame and keypoints are therefore only detected in the sequential frame. They are detected , matched and reconstructed in the same fashion as the first pair of frames, however to ensure that the reconstructed points are in the same scale as the points reconstructed from the first pair of frames the motion of the camera between t he frames is estimated from 3D-2D correspondences using a best practice algorithm. If the purpose of progressive reconstruction is for visualization the best practice combination algorithm for keypoint detection was found to be Speeded Up Robust Features (SURF) as it results in more reconstructed points than Scale-Invariant Feature Transform (SIFT). SIFT is however more computationally efficient and thus better suited if the number of reconstructed points does not matter, for example if the purpose of progressive reconstruction is for camera tracking. For all purposes the best practice combination algorithm for matching was found to be optical flow as it is the most efficient and for ego-motion estimation the best practice combination algorithm was found to be the 5-point algorithm as it is robust to points located on planes. This research is significant as the effects of the key steps of progressive reconstruction and the choices made at each step on the accuracy and efficiency of the reconstruction as a whole have never been studied. As a result progressive stereo reconstruction can now be performed in near real-time on a mobile device without compromising the accuracy of reconstruction. |
| format | Thesis |
| id | oai:open.uct.ac.za:11427/24320 |
| institution | University of Cape Town (South Africa) |
| language | eng |
| last_indexed | 2026-06-10T12:33:41.762Z |
| license_str | Not specified — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from UCTD — University of Cape Town Open Access Repository |
| publishDate | 2017 |
| publishDateRange | 2017 |
| publishDateSort | 2017 |
| publisher | School of Architecture, Planning and Geomatics |
| publisherStr | School of Architecture, Planning and Geomatics |
| record_format | dspace |
| source_str | UCTD — University of Cape Town Open Access Repository |
| spelling | oai:open.uct.ac.za:11427/24320 An empirical assessment of real-time progressive stereo reconstruction Westaway, Matthew Sithole, George Geomatics 3D reconstruction from images, the problem of reconstructing depth from images, is one of the most well-studied problems within computer vision. In part because it is academically interesting, but also because of the significant growth in the use of 3D models. This growth can be attributed to the development of augmented reality, 3D printing and indoor mapping. Progressive stereo reconstruction is the sequential application of stereo reconstructions to reconstruct a scene. To achieve a reliable progressive stereo reconstruction a combination of best practice algorithms needs to be used. The purpose of this research is to determine the combinat ion of best practice algorithms that lead to the most accurate and efficient progressive stereo reconstruction i.e the best practice combination. In order to obtain a similarity reconstruction the in t rinsic parameters of the camera need to be known. If they are not known they are determined by capturing ten images of a checkerboard with a known calibration pattern from different angles and using the moving plane algori thm. Thereafter in order to perform a near real-time reconstruction frames are acquired and reconstructed simultaneously. For the first pair of frames keypoints are detected and matched using a best practice keypoint detection and matching algorithm. The motion of the camera between the frames is then determined by decomposing the essential matrix which is determined from the fundamental matrix, which is determined using a best practice ego-motion estimation algorithm. Finally the keypoints are reconstructed using a best practice reconstruction algorithm. For sequential frames each frame is paired with t he previous frame and keypoints are therefore only detected in the sequential frame. They are detected , matched and reconstructed in the same fashion as the first pair of frames, however to ensure that the reconstructed points are in the same scale as the points reconstructed from the first pair of frames the motion of the camera between t he frames is estimated from 3D-2D correspondences using a best practice algorithm. If the purpose of progressive reconstruction is for visualization the best practice combination algorithm for keypoint detection was found to be Speeded Up Robust Features (SURF) as it results in more reconstructed points than Scale-Invariant Feature Transform (SIFT). SIFT is however more computationally efficient and thus better suited if the number of reconstructed points does not matter, for example if the purpose of progressive reconstruction is for camera tracking. For all purposes the best practice combination algorithm for matching was found to be optical flow as it is the most efficient and for ego-motion estimation the best practice combination algorithm was found to be the 5-point algorithm as it is robust to points located on planes. This research is significant as the effects of the key steps of progressive reconstruction and the choices made at each step on the accuracy and efficiency of the reconstruction as a whole have never been studied. As a result progressive stereo reconstruction can now be performed in near real-time on a mobile device without compromising the accuracy of reconstruction. 2017-05-16T08:00:25Z 2017-05-16T08:00:25Z 2016 Master Thesis Masters MEng http://hdl.handle.net/11427/24320 eng application/pdf School of Architecture, Planning and Geomatics Faculty of Engineering and the Built Environment University of Cape Town |
| spellingShingle | Geomatics Westaway, Matthew An empirical assessment of real-time progressive stereo reconstruction |
| thesis_degree_str | Master's |
| title | An empirical assessment of real-time progressive stereo reconstruction |
| title_full | An empirical assessment of real-time progressive stereo reconstruction |
| title_fullStr | An empirical assessment of real-time progressive stereo reconstruction |
| title_full_unstemmed | An empirical assessment of real-time progressive stereo reconstruction |
| title_short | An empirical assessment of real-time progressive stereo reconstruction |
| title_sort | empirical assessment of real time progressive stereo reconstruction |
| topic | Geomatics |
| url | http://hdl.handle.net/11427/24320 |
| work_keys_str_mv | AT westawaymatthew anempiricalassessmentofrealtimeprogressivestereoreconstruction AT westawaymatthew empiricalassessmentofrealtimeprogressivestereoreconstruction |