Weakly supervised learning for compressed video analysis on retrieval and classification tasks for visual alert

Abstract

Several machine learning techniques have relied on large labeled data sets to construct predictive models and solving supervised learning tasks. The use of deep learning techniques can be highlighted, since it have been broadly and successfully used in various domains. On the other hand, in many circumstances, the labeled sets are unavailable or insufficient to train effective supervised models. Such scenarios have been mainly addressed by unsupervised learning techniques, which consider the unlabeled data to learn about its structure. However, the use of completely unsupervised methods still remains a research challenge in many scenarios and situations. A promising solution is based on the use of weakly supervised approaches, capable of performing effective learning tasks based on incomplete or inaccurate labeled sets. In this project, we intend to investigate the analysis, retrieval, and classification of compressed video domain based on small training sets. The main object of the project consists in to investigate and propose methods capable of analyzing compressed video sequences and trigger alerts according to considered applications. Such approaches can be useful and relevant in several domains, ranging from surveillance, medical and industrial environments to smart homes. The fundamental research challenge consists in making use of different techniques in order to analyze, represent, and classification videos using restricted labeled data. The proposed approach aims at exploiting the maximum available information, in order to become the approach suitable for operating with small training datasets. We intend to exploit: (i) deep learning representations; (ii) contextual unsupervised measures and; (iii) fusion techniques, in order to extend the initial labeled sets. The first challenge to be addressed is to analyze and represent videos in the compressed domain using deep learning techniques. Based on such representations, we intend to investigate strategies for expanding the training sets using unsupervised contextual measures. Given the obtained labeled sets, fusion strategies will be used to combined diverse classification methods and triggering alerts. Although the methods which will investigated can be used in several domains, we intend to select domains to validate the proposed approaches. The selection will be performed considering the existence of public available datasets to conduct experimental evaluations. (AU)