. "http://www.isvav.cz/projectDetail.do?rowId=LL1303"^^ . "2015-02-16+01:00"^^ . "category localization large scale image collections"@en . . "2013-07-01+02:00"^^ . "0"^^ . . "2014-01-31+01:00"^^ . . . "The main goal of the project is to scale the category-object retrieval to large databases. To achieve this, a suitable image representation for indexing will be built. To this end, the following will be investigated: Design of novel local features and feature classes, and suitable descriptors that enable both fast indexing and have discriminative power for verification. Machine learning approaches to learn features and descriptors from a large amounts of training data. The data will be obtained in an unsupervised manner through large-scale mining The following issues will be investigated for the category model learning and detection stages: Category models capable of capturing intra-class variation will be studied. Methods for a transition from generative to discriminative models will be investigated. Algorithms for fast scoring eliminating a majority of false candidates and preserving most of the instances of the category will be investigated. We will pursue hypothesize and verify algorithms with sequential verification focusing on the speed v. quality of the decision trade-off. Special attention will be paid to spatial or topological distribution of object parts. On the high-level side we plan to tackle the following: Performance boosting by incremental model update from novel examples detected from unlabelled data. Mining for commonly appearing category-like structures."@en . . . "Hlavn\u00EDm c\u00EDlem projektu je vyhled\u00E1v\u00E1n\u00ED kategori\u00ED objekt\u016F ve velk\u00FDch obrazov\u00FDch datab\u00E1z\u00EDch. K dosa\u017En\u00ED tohoto c\u00EDle vyvineme reprezentaci obr\u00E1zk\u016F vhodnou pro indexov\u00E1n\u00ED. N\u00E1sleduj\u00EDc\u00ED podprobl\u00E9my budou zkoum\u00E1ny: n\u00E1vrh nov\u00FDch typ\u016F lok\u00E1ln\u00EDch oblast\u00ED z\u00E1jmu a vhodn\u00FDch deskriptor\u016F, kter\u00E9 budou vhodn\u00E9 pro indexov\u00E1n\u00ED a z\u00E1rove\u0148 budou dostate\u010Dn\u011B diskriminativn\u00ED pro ov\u011B\u0159ov\u00E1n\u00ED. Metody strojov\u00E9ho u\u010Den\u00ED budou aplikov\u00E1ny k nau\u010Den\u00ED detektor\u016F bod\u016F z\u00E1jmu a deskriptor\u016F z velk\u00E9ho mno\u017Estv\u00ED tr\u00E9novac\u00EDch dat. Tr\u00E9novac\u00ED data budou z\u00EDsk\u00E1na automaticky za pomoci vyt\u011B\u017Eov\u00E1n\u00ED z velk\u00FDch kolekc\u00ED dat. P\u0159i u\u010Den\u00ED modelu kategori\u00ED a p\u0159i detekci model\u016F budeme studovat: modely zachycuj\u00EDc\u00ED variaci v r\u00E1mci jedn\u00E9 t\u0159\u00EDdy a metody p\u0159echodu od generativn\u00EDch model\u016F k model\u016Fm diskriminativn\u00EDm. Algoritmy pro rychl\u00E9 sk\u00F3rov\u00E1n\u00ED, kter\u00E9 eliminuj\u00ED v\u011Bt\u0161inu chybn\u00FDch kandid\u00E1t\u016F a sou\u010Dasn\u011B zachovaj\u00ED v\u011Bt\u0161inu instanc\u00ED objekt\u016F hledan\u00E9 kategorie budou studov\u00E1ny. Budeme vyv\u00EDjet algoritmy typu hypot\u00E9za a verifikace se sekven\u010Dn\u00EDm rozhodov\u00E1n\u00EDm s d\u016Frazem na vztah mezi rychlost\u00ED a kvalitou rozhodnut\u00ED. Zvl\u00E1\u0161tn\u00ED pozornost bude v\u011Bnov\u00E1na verifikaci pomoc\u00ED geometrick\u00E9 a topologick\u00E9 distribuce oblast\u00ED z\u00E1jmu. Na vy\u0161\u0161\u00ED \u00FArovni abstrakce pl\u00E1nujeme zkoumat zvy\u0161ov\u00E1n\u00ED kvality vyhled\u00E1v\u00E1n\u00ED pomoc\u00ED inkrement\u00E1ln\u00EDho vylep\u0161ov\u00E1n\u00ED modelu z nov\u00FDch p\u0159\u00EDklad\u016F nalezen\u00FDch v neozna\u010Den\u00FDch datech, a vyt\u011B\u017Eov\u00E1n\u00ED \u010Dasto se vyskytuj\u00EDc\u00EDch struktur reprezentuj\u00EDc\u00EDch kategorii objekt\u016F." . . . . "6"^^ . . . . "Vyhled\u00E1v\u00E1n\u00ED vizu\u00E1ln\u00EDch kategori\u00ED ve velk\u00E9m mno\u017Estv\u00ED obr\u00E1zk\u016F" . "6"^^ . . "LL1303" . "0"^^ . . "Large Scale Category Retrieval"@en . "1"^^ . . . . "2018-06-30+02:00"^^ .