Home |  ENGLISH |  Kontakt |  Impressum |  Anmelden |  KIT

Christian Hentschel

Aus Aifbportal

Wechseln zu: Navigation, Suche

Training Visual Concept Classifiers


Visual Concept Detection describes the process of automatically classifying images and video based on the depicted visual content. This talk will start by comparing different approaches for vis-ual concept detection, namely Bag-of-Visual-Words and Deep Convolutional Neural Networks (CNN). Bag-of-Visual-Words methods represented the de facto standard until CNNs emerged, backed by highly parallel hardware as well as large training datasets.

The talk will present the impact of the available amount of training data on the classification perfor-mance as achieved by the individual approaches. Furthermore, techniques for model visualization will be presented. Non-linear models suffer from the lack of interpretability. The presented visualiza-tion methods help to qualitatively compare visual concept models by highlighting image regions considered important for the final classification decision.

Finally, the talk will address the problem of leveraging social photo communities in order to increase the amount of available training data without additional manual labeling efforts. A social community language model will be presented as well as an outlook for multi-modal retrieval.

(Christian Hentschel)

Start: 12. Januar 2018 um 14:00
Ende: 12. Januar 2018 um 15:00

Im Gebäude 05.20, Raum: 1C-04

Veranstaltung vormerken: (iCal)

Veranstalter: Forschungsgruppe(n) Information Service Engineering
Information: Media:Hentschel 12-01-2018.pdf