Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Computing | en_US |
| dc.creator | Ji, Peng | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/11372 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Multi-level structured output space domain adaptation for semantic segmentation | en_US |
| dcterms.abstract | Using domain adaptation is essential for semantic image segmentation, because it is costly and time-consuming to manually label large data sets with pixel-level tags. It is of great interest to develop algorithms that enable source ground truth labels to be compatible with target domain. As source domain has spatial similarities with the target one in terms of structured outputs of semantic segmentation, we propose to employ adversarial learning in the aspect of output space so as to produce confident pseudo label to close the gap between source domain and target one. To further enhance adapted model at different feature levels, a multi-level adversarial network is constructed by us for efficient output space domain adaptation. In addition, in the scene segmentation task, similar features would be connected with each other in spite of the distance between them and different semantic receptions are also interrelated. Feature representation makes contribution to more accurate domain adaptation. In order to further improve it, we don't forward the input images to the segmentation network until perform a scene segmentation task based on self-attentive mechanism by catching rich context dependencies. In particular, we attach two types of attention modules to the extended Convolution Neural Networks before inputting the images into adaption segmentation network. Moreover, we propose a not complicated and useful pseudo label selection strategy to generate a trusted pseudo-label for the target instance, which bridges the gap between source domain and target one in the aspect of distribution. A lot of experiments and studies of ablation are performed with various field adaptation settings, including "GTA to Cityscapes" and "SYNTHIA to Cityscapes". It is showed that the proposed method is on good form compared to the state-of-the-art methods in respect to precision and visual quality. | en_US |
| dcterms.extent | 4, iv, 40 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2021 | en_US |
| dcterms.educationalLevel | M.Sc. | en_US |
| dcterms.educationalLevel | All Master | en_US |
| dcterms.LCSH | Machine learning | en_US |
| dcterms.LCSH | Computer algorithms | en_US |
| dcterms.LCSH | Hong Kong Polytechnic University -- Dissertations | en_US |
| dcterms.accessRights | restricted access | en_US |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| 5820.pdf | For All Users (off-campus access for PolyU Staff & Students only) | 1.28 MB | Adobe PDF | View/Open |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/11372