Schema för examensarbete presentationer vid Institutionen för Datavetenskap, den 31:e oktober.
Kl 13.15: Cloud Based Oﬄoading of Multimedia Tasks
Författare/Authors: Hani Fakhouri, Michal Sadowski
Handledare/Supervisor: Sunil Shah (Sony Mobile Communications)
Examinator/Examiner: Patrik Persson (LTH)
Computation offloading into cloud based servers has become more and more popular through the last decade. In this thesis offloading of video transcoding was investigated to determine the most suitable transcoding parameters, when it is feasible to offload video for transcoding and what can be gained from offloading. For this purpose we develop an offloading framework making it possible to upload a media file to the cloud, transcode it and then download the transcoded file. This framework is the base for experiments investigating the transcoding parameters, battery consumption and comparing media file qualities. The framework design and ability of future extension is shown together with the experimental results. The results show that the bit rate should not be decreased more then 70-80% due to noticeable quality degradation. Results also show that the power consumption and execution time are lower when using the offloading framework when the connection bandwidth is higher than 6-7Mbps. To conclude, offloading of media files for transcoding gives more flexibility and can be used to decrease the file size of media files while keeping the similar quality.
Kl 14.15: Identifying Entities across Images and Text
Författare/Authors: Rebecka Weegar
Handledare/Supervisor: Pierre Nugues (LTH)
Examinator/Examiner: Jacek Malec (LTH)
Given an image and a caption describing that image, I present methods to find a mapping between entities in the image and the mentions of the same entities in the caption. To investigate this, I used a large data set of images with short captions, where the regions in the images are manually segmented and labeled. To find a mapping, interesting entities need to be extracted from the captions before they can be matched with the labeled segments. The mapping is then done by using semantic similarity between the words describing objects in text and image and by using statistical information extracted from the data set. As a final step, I used a reranker. The reranker uses machine learning techniques to improve the mapping. Features for the reranker includes syntactic relationships between the words in the captions, and spatial relationships between segments in the picture. Using these techniques, I could match objects in pictures to their correct mentions for 89 percent of segments, when such a matching exists.
Vill du opponera?
Anmäl dig som opponent till den examinator som finns angiven vid den presentation som du är intresserad av, maila till Firstname.Lastname@cs.lth.se. Antagning sker i mån av plats. Ta med dig en skriftlig opponering till presentationen, att dela med examinatorn och den som presenterar. Oppositionen examineras individuellt.
Presentationerna äger rum i E-huset, Ole Römersväg 3. Lokal: E:2116