EVENTO
DEEP VIDEO SUMMARIZATION: BUILDING STRATEGIES WITH A DEEP LEARNING-BASED APPROACH TO CONSTRUCT VIDEO SUMMARIES
Tipo de evento: Exame de Qualificação
A video summary, containing important information along the video captured may reduce the effort in a social media query, or for a surveillance or traffic analysis. Video summarization approaches generally are based in machine learning techniques, using Deep Learning in the supervised methodologies and mainly clustering in the unsupervised ones [1]. For the supervised approaches, there is a ground truth, content with labels associated with it. However, a ground truth is not always available, and an unsupervised methodology is employed to deal with this drawback [2]. An alternative solution little-used in video summarization purposes is the semi-supervised approach, where data with a ground truth and data without it, is used to yield the summaries [3].Given this, we present in [4] a pipeline for video summarization, employing a video query with the correspondent ground truth; a semi-supervised autoencoder trained with sample for the video query and from videos without ground truth; and, clustering with k-means to select the key-frames to compose the summary. The present proposal aims to enhance this pipeline applying machine learning techniques, mainly Deep Learning ones. It includes: to produce summaries with a semantic approach [1] and also more close to the user preferences, using query with text; to employ a methodology to automatically determine the autoencoder architecture using genetic algorithms; to apply Pre-trained neural networks [5] instead of autoencoder; clustering with Deep Learning. Furthermore, the proposal also includes the design of a measure to evaluate the summaries, that will be based in the importance categories: representativeness (key-frames that represent the entire class) and diversity (key-frames that express the amount of characteristics in the video) [1]; and the adaptation of the pipeline for summarization in fluid dynamic simulation context [6].REFERÊNCIAS:[1] M. Basavarajaiah and P. Sharma, Survey of compressed domain video summarization techniques,ACM Comput. Surv., vol. 52, no. 6, Oct. 2019.[2] L. Yuan, F. E. Tay, P. Li, L. Zhou, and J. Feng, Cycle-sum: cycle-consistent adversarial lstmnetworks for unsupervised video summarization, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, 2019, pp. 91439150.[3] T.-J. Fu, S.-H. Tai, and H.-T. Chen, Attentive and adversarial learning for video summariza-tion, in IEEE Winter Conference on Applications of Computer Vision (WACV), 2019.[4] E. Silva, E. M. Ramos., L. T. da Silva., J. S. Cardoso., and G. A. Giraldi., Video summariza-tion through total variation, deep semi-supervised autoencoder and clustering algorithms, in Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 4: VISAPP,, INSTICC.SciTePress,2020, pp. 315322[5] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press, 2016, http://www.deeplearningbook.org.[6] E. Ramos, L. da Silva, J. S. Cardoso, and G. Giraldi, Deep neural network for vectorfield topology recognition with applications to fluid flow summarization,XXXIX Ibero-LatinAmerican Congress on Computational Methods in Engineering (CILAMCE), 2018.Para assistir essa defesa acesse : https://us02web.zoom.us/j/85400127286?pwd=cWp6ZHNIK2RsWjViTGN1cG92SWYrdz09
Data Início: 27/08/2020 Hora: 14:00 Data Fim: 27/08/2020 Hora: 16:00
Local: LNCC - Laboratório Nacional de Computação Ciêntifica - Webinar
Aluno: Éden Pereira da Silva - LNCC -
Orientador: Gilson Antônio Giraldi - Laboratório Nacional de Computação Científica - LNCC
Participante Banca Examinadora: Evandro Ottoni Teatini Salles - - UFES Fabio Andre Machado Porto - Laboratório Nacional de Computação Científica - LNCC Fernando Von Zuben - - UNICAMP Jack Baczynski - Laboratório Nacional de Computação Científica - LNCC
Suplente Banca Examinadora: Artur Ziviani - Laboratório Nacional de Computação Científica - LNCC
