Logo do repositório

Percorrer por autor "Tsironi, Eleni"

A mostrar 1 - 1 de 1
  • Image summarisation: human action description from static images
    Publication . Tsironi, Eleni; Baptista, Jorge Manuel Evangelista; Madec, Henri; Orăsan, Constantin
    The object of this master thesis is Image Summarisation and more specifically the automatic human action description from static images. The work has been organised into three main phases, with first one being the data collection, second the actual system implementation and third the system evaluation. The dataset consists of 1287 images depicting human activities belonging in fours semantic categories; "walking a dog", "riding a bike", "riding a horse" and "playing the guitar". The images were manually annotated with an approach based in the idea of crowd sourcing, and the annotation of each sentence is in the form of one or two simple sentences. The system is composed by two parts, a Content-based Image Retrieval part and a Natural Language Processing part. Given a query image the first part retrieves a set of images perceived as visually similar and the second part processes the annotations following each of the images in order to extract common information by using a graph merging technique of the dependency graphs of the annotated sentences. An optimal path consisting of a subject-verb-complement relation is extracted and transformed into a proper sentence by applying a set of surface processing rules. The evaluation of the system was carried out in three different ways. Firstly, the Content-based Image Retrieval sub-system was evaluated in terms of precision and recall and compared to a baseline classification system based on randomness. In order to evaluate the Natural Language Processing sub-system, the Image Summarisation task was considered as a machine translation task, and therefore it was evaluated in terms of BLEU score. Given images that correspond to the same semantic as a query image the system output was compared to the corresponding reference summary as provided during the annotation phase, in terms of BLEU score. Finally, the whole system has been qualitatively evaluated by means of a questionnaire. The conclusions reached by the evaluation is that even if the system does not always capture the right human action and subjects and objects involved in it, it produces understandable and efficient in terms of language summaries.
    2014Dissertação de mestrado Acesso aberto Ver mais