|
|
Visipedia - Embedding-driven Visual Feature Extraction and Learning
Jakeš, Jan ; Beran, Vítězslav (oponent) ; Zemčík, Pavel (vedoucí práce)
Multidimensional embedding is a powerful method of representing similarity measures among objects without the need for their explicit categorization. It has been increasingly used in recent years to annotate objects making an important part of the Visipedia project and its related work. This work explores the possibilities of learning from embedding-annotated images using their visual attributes and develops methods of predicting embedding coordinates for previously unseen images. It studies the relevant feature extraction and learning algorithms and describes the whole process of design and development of such a system using common machine learning approaches. The system is tested and evaluated with two different datasets and the performed experiments present the first results for a task of its kind.
|
|
|
Visipedia - Multi-Dimensional Object Embedding Based on Perceptual Similarity
Matera, Tomáš ; Hradiš, Michal (oponent) ; Zemčík, Pavel (vedoucí práce)
Some problems like fine-grained categorization or human-based computation has become popular in recent years in the community, which has been proven by a large number of published works concerning these topics. Whereas most of these works uses a "classical'' visual features extracted by machine, this one in particular focuses on perceptual properties which cannot be easily sampled by machine and which involves humans into this data retrieval process. There are examined ways, how to obtain perceptual similarities from humans cheaply and effectively also in terms of scalability. There are performed various experiments and purposed several methods to improve this efficiency. The work also reviews and compares existing methods of embedding learning and navigating through its space. The acquired observations are subsequently used in a complex experiment evaluated with a food image dataset, covering the whole procedure from similarity retrieval from humans, over data embedding learning up to searching in such multi-dimensional space.
|
|
|
Visipedia - Embedding-driven Visual Feature Extraction and Learning
Jakeš, Jan ; Beran, Vítězslav (oponent) ; Zemčík, Pavel (vedoucí práce)
Multidimensional embedding is a powerful method of representing similarity measures among objects without the need for their explicit categorization. It has been increasingly used in recent years to annotate objects making an important part of the Visipedia project and its related work. This work explores the possibilities of learning from embedding-annotated images using their visual attributes and develops methods of predicting embedding coordinates for previously unseen images. It studies the relevant feature extraction and learning algorithms and describes the whole process of design and development of such a system using common machine learning approaches. The system is tested and evaluated with two different datasets and the performed experiments present the first results for a task of its kind.
|
|
|
Visipedia - Multi-Dimensional Object Embedding Based on Perceptual Similarity
Matera, Tomáš ; Hradiš, Michal (oponent) ; Zemčík, Pavel (vedoucí práce)
Some problems like fine-grained categorization or human-based computation has become popular in recent years in the community, which has been proven by a large number of published works concerning these topics. Whereas most of these works uses a "classical'' visual features extracted by machine, this one in particular focuses on perceptual properties which cannot be easily sampled by machine and which involves humans into this data retrieval process. There are examined ways, how to obtain perceptual similarities from humans cheaply and effectively also in terms of scalability. There are performed various experiments and purposed several methods to improve this efficiency. The work also reviews and compares existing methods of embedding learning and navigating through its space. The acquired observations are subsequently used in a complex experiment evaluated with a food image dataset, covering the whole procedure from similarity retrieval from humans, over data embedding learning up to searching in such multi-dimensional space.
|
host ::
RSS.