Large-scale Multi-label Video Classification
Year: 2018
Type: dissertation
Abstract: Deep learning has been established as a powerful method to tacklevideo classification tasks. Encouraged by the remarkable success, thisthesis propose several effective classification models rely on the largestvideo dataset YouTube-8M. First, we investigate the diversity of theDense Neural Nets (DNN) by exploring several factors which controlthe corresponding performance. Furthermore, inspired by the wideand deep classification model which was originally used in the recommendationsystem, we build up a new classification model whichmodify the wide and deep part separately. Despite that, the thesis alsoprovides extensive experiments and comparisons on the performanceof proposed models. Experimental results find out the audio featureis able to bring valuable help to the final performance. The result hasshown that both models outperform the benchmark, although there isnarrow gap with the state-of-the-art. We also evaluate the efficiency ofmodel on real-world datasets, the positive outcome make it possiblefor production in the future.
Author: Jinzhou Cai
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2891189600