{"id":{"repo_id":"birmingham","oai_identifier":"oai:etheses.bham.ac.uk:7954"},"canonical_url":"https://search.ndltd.org/etd/birmingham/oai:etheses.bham.ac.uk:7954","repository":{"repo_id":"birmingham","name":"University of Birmingham","base_url":"https://etheses.bham.ac.uk/cgi/oai2"},"display":{"title":"3D compositional hierarchies for object categorization","abstract":"Deep learning methods have become the default tool for image classification. However, application of deep learning to surface shape classification is burdened by the limitations of existing methods, in particular, by lack of invariance to geometric transformations of input data. This thesis proposes two novel frameworks for learning a multi-layer representation of surface shape features, namely the view-based and the surface-based compositional hierarchical frameworks. The proposed representation is a hierarchical vocabulary of shape features, termed parts. Parts of the first layer are pre-defined, while parts of the subsequent layers, describing spatial relations of subparts, are learned. The view-based framework describes spatial relations between subparts using a camera-based reference frame. The key stage of the learning algorithm is part selection which forms the vocabulary based on multi-objective optimization, considering different importance measures of parts. Our experiments show that this framework enables efficient category recognition on a large-scale dataset. The surface-based framework exploits part-based intrinsic reference frames, which are computed for lower layers parts and inherited by parts of the subsequent layers. During learning spatial relations between subparts are described in these reference frames. During inference, a part is detected in input data when its subparts are detected at certain positions and orientations in each other’s reference frames. Since rigid body transformations don’t change positions and orientations of parts in intrinsic reference frames, this approach enables efficient recognition from unseen poses. Experiments show that this framework exhibits a large discriminative power and greater robustness to rigid body transformations than advanced CNN-based methods.","abstract_html":"Deep learning methods have become the default tool for image classification. However, application of deep learning to surface shape classification is burdened by the limitations of existing methods, in particular, by lack of invariance to geometric transformations of input data. This thesis proposes two novel frameworks for learning a multi-layer representation of surface shape features, namely the view-based and the surface-based compositional hierarchical frameworks. The proposed representation is a hierarchical vocabulary of shape features, termed parts. Parts of the first layer are pre-defined, while parts of the subsequent layers, describing spatial relations of subparts, are learned. The view-based framework describes spatial relations between subparts using a camera-based reference frame. The key stage of the learning algorithm is part selection which forms the vocabulary based on multi-objective optimization, considering different importance measures of parts. Our experiments show that this framework enables efficient category recognition on a large-scale dataset. The surface-based framework exploits part-based intrinsic reference frames, which are computed for lower layers parts and inherited by parts of the subsequent layers. During learning spatial relations between subparts are described in these reference frames. During inference, a part is detected in input data when its subparts are detected at certain positions and orientations in each other’s reference frames. Since rigid body transformations don’t change positions and orientations of parts in intrinsic reference frames, this approach enables efficient recognition from unseen poses. Experiments show that this framework exhibits a large discriminative power and greater robustness to rigid body transformations than advanced CNN-based methods.","abstract_has_math":false,"creators":["Kramarev, Vladislav Vadimovich"],"institution":"University of Birmingham","degree_name":"d_ph","degree_level":"d_ph","degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2017,"date_issued":"2017-12","date_published":"2017-12","updated_at":"2026-11-19T09:29:41Z","subjects":["TK Electrical engineering. Electronics Nuclear engineering"],"languages":["eng"],"rights":[],"rights_urls":[],"identifier_entries":[],"person_identifier_entries":[]},"links":{"outbound_url":null,"outbound_label":null,"outbound_source":null,"record_url":null,"record_label":null,"record_kind":null,"record_source":null,"download_url":"http://etheses.bham.ac.uk//id/eprint/7954/1/Kramarev17PhD.pdf","download_label":"Download PDF","download_source":"dc:identifier.uri","access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://etheses.bham.ac.uk/cgi/oai2?verb=GetRecord&metadataPrefix=uketd_dc&identifier=oai%3Aetheses.bham.ac.uk%3A7954","prefix":"uketd_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.sponsor","label":"Sponsor","values":["na"]},{"key":"dc:creator","label":"Author","values":["Kramarev, Vladislav Vadimovich"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2017-12"]},{"key":"dc:date.issued","label":"Date","values":["2017-12"]},{"key":"dc:publisher.department","label":"Dc Publisher Department","values":["College of Engineering & Physical Sciences","School of Computer Science"]},{"key":"dc:publisher.institution","label":"Dc Publisher Institution","values":["University of Birmingham"]},{"key":"dc:relation.isreferencedby","label":"Dc Relation Isreferencedby","values":["http://etheses.bham.ac.uk//id/eprint/7954/"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"dc:type.qualificationlevel","label":"Dc Type Qualificationlevel","values":["d_ph"]},{"key":"dc:type.qualificationname","label":"Dc Type Qualificationname","values":["d_ph"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["TK Electrical engineering. Electronics Nuclear engineering"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.dcterms_iso639_2","label":"Language (ISO 639-2)","values":["eng"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["http://etheses.bham.ac.uk//id/eprint/7954/1/Kramarev17PhD.pdf","http://etheses.bham.ac.uk//id/eprint/7954/2/Decl_IS_Kramarev17PhD.pdf"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Deep learning methods have become the default tool for image classification. However, application of deep learning to surface shape classification is burdened by the limitations of existing methods, in particular, by lack of invariance to geometric transformations of input data. This thesis proposes two novel frameworks for learning a multi-layer representation of surface shape features, namely the view-based and the surface-based compositional hierarchical frameworks. The proposed representation is a hierarchical vocabulary of shape features, termed parts. Parts of the first layer are pre-defined, while parts of the subsequent layers, describing spatial relations of subparts, are learned. The view-based framework describes spatial relations between subparts using a camera-based reference frame. The key stage of the learning algorithm is part selection which forms the vocabulary based on multi-objective optimization, considering different importance measures of parts. Our experiments show that this framework enables efficient category recognition on a large-scale dataset. The surface-based framework exploits part-based intrinsic reference frames, which are computed for lower layers parts and inherited by parts of the subsequent layers. During learning spatial relations between subparts are described in these reference frames. During inference, a part is detected in input data when its subparts are detected at certain positions and orientations in each other’s reference frames. Since rigid body transformations don’t change positions and orientations of parts in intrinsic reference frames, this approach enables efficient recognition from unseen poses. Experiments show that this framework exhibits a large discriminative power and greater robustness to rigid body transformations than advanced CNN-based methods."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dcterms:accessrights","label":"Dcterms Accessrights","values":["public","staffonly"]},{"key":"dc:title","label":"Title","values":["3D compositional hierarchies for object categorization"]}]}],"canonical_facts":{"dc:contributor.sponsor":["na"],"dc:creator":["Kramarev, Vladislav Vadimovich"],"dc:date":["2017-12"],"dc:date.issued":["2017-12"],"dc:description.abstract":["Deep learning methods have become the default tool for image classification. However, application of deep learning to surface shape classification is burdened by the limitations of existing methods, in particular, by lack of invariance to geometric transformations of input data. This thesis proposes two novel frameworks for learning a multi-layer representation of surface shape features, namely the view-based and the surface-based compositional hierarchical frameworks. The proposed representation is a hierarchical vocabulary of shape features, termed parts. Parts of the first layer are pre-defined, while parts of the subsequent layers, describing spatial relations of subparts, are learned. The view-based framework describes spatial relations between subparts using a camera-based reference frame. The key stage of the learning algorithm is part selection which forms the vocabulary based on multi-objective optimization, considering different importance measures of parts. Our experiments show that this framework enables efficient category recognition on a large-scale dataset. The surface-based framework exploits part-based intrinsic reference frames, which are computed for lower layers parts and inherited by parts of the subsequent layers. During learning spatial relations between subparts are described in these reference frames. During inference, a part is detected in input data when its subparts are detected at certain positions and orientations in each other’s reference frames. Since rigid body transformations don’t change positions and orientations of parts in intrinsic reference frames, this approach enables efficient recognition from unseen poses. Experiments show that this framework exhibits a large discriminative power and greater robustness to rigid body transformations than advanced CNN-based methods."],"dc:format":["application/pdf"],"dc:identifier.uri":["http://etheses.bham.ac.uk//id/eprint/7954/1/Kramarev17PhD.pdf","http://etheses.bham.ac.uk//id/eprint/7954/2/Decl_IS_Kramarev17PhD.pdf"],"dc:language.dcterms_iso639_2":["eng"],"dc:publisher.department":["College of Engineering & Physical Sciences","School of Computer Science"],"dc:publisher.institution":["University of Birmingham"],"dc:relation.isreferencedby":["http://etheses.bham.ac.uk//id/eprint/7954/"],"dc:subject":["TK Electrical engineering. Electronics Nuclear engineering"],"dcterms:accessrights":["public","staffonly"],"dc:title":["3D compositional hierarchies for object categorization"],"dc:type":["Thesis"],"dc:type.qualificationlevel":["d_ph"],"dc:type.qualificationname":["d_ph"]},"updated_at":"2026-11-19T09:29:41Z"}