Unsupervised Domain Adaptation for Object Counting
| dc.contributor.advisor | Stavness, Ian | |
| dc.contributor.committeeMember | Stanley, Kevin | |
| dc.contributor.committeeMember | Roy, Chanchal | |
| dc.contributor.committeeMember | Bui, Francis | |
| dc.creator | Ayalew, Tewodros Wondifraw | |
| dc.date.accessioned | 2021-01-27T17:41:35Z | |
| dc.date.available | 2021-01-27T17:41:35Z | |
| dc.date.created | 2020-11 | |
| dc.date.issued | 2021-01-27 | |
| dc.date.submitted | November 2020 | |
| dc.date.updated | 2021-01-27T17:41:35Z | |
| dc.description.abstract | Supervised learning is a common approach for counting objects in images, but for counting small, densely located objects, the required image annotations are burdensome to collect. Counting plant organs for image-based plant phenotyping and crowd counting fall within this category. Object counting in plant images is further challenged by having plant image datasets with significant domain shift due to different experimental conditions, e.g. applying an annotated dataset of indoor plant images for use on outdoor images, or on a different plant species. Learning to count from synthetic data to apply the knowledge in real-world data is also another important domain shift addressed in crowd counting tasks where getting annotations for real-world images, especially for highly crowded images, is tedious and potentially inaccurate. In this thesis, we propose a domain-adversarial learning approach for domain adaptation of density map estimation for the purposes of object counting. We took a fully convolutional network --- initially designed for image segmentation --- and trained it to count objects via density estimation from images sampled from a distribution and in parallel adapted the knowledge to a related counting task without the need of annotations. The proposed approach does not assume perfectly aligned distributions between the source and target datasets, which makes it more broadly applicable within general object counting and plant organ counting tasks. Evaluation on three diverse object counting tasks (wheat spikelets, leaves, crowd) demonstrate consistent performance on the target datasets across different classes of domain shift: from indoor-to-outdoor images, from species-to-species, and from synthetic-to-real. | |
| dc.format.mimetype | application/pdf | |
| dc.identifier.uri | http://hdl.handle.net/10388/13235 | |
| dc.subject | Object counting | |
| dc.subject | Unsupervised Domain Adaptation | |
| dc.subject | Density-based counting | |
| dc.subject | Deep learning | |
| dc.subject | Transfer learning | |
| dc.subject | indoor-to-outdoor domain adaptation | |
| dc.subject | species-to-species domain adaptation | |
| dc.subject | synthetic-to-real domain adaptation | |
| dc.title | Unsupervised Domain Adaptation for Object Counting | |
| dc.type | Thesis | |
| dc.type.material | text | |
| thesis.degree.department | Computer Science | |
| thesis.degree.discipline | Computer Science | |
| thesis.degree.grantor | University of Saskatchewan | |
| thesis.degree.level | Masters | |
| thesis.degree.name | Master of Science (M.Sc.) |