• Re-Ranking and Rank Aggregation Approaches for Image Retrieval Tasks.
  Foundation for Research Support of the State of São Paulo (Fundação de Àmparo a Pesquisa do Estado de São Paulo - FAPESP) - Grant 2013/08645-0
  
  2014-2018
  Modality Young Researcher Program, Principal Investigator: Daniel Carlos Guimarães Pedronette.

Abstract:
Content-Based Image Retrieval (CBIR) systems aims at retrieving the most similar images in a collection by taking into account image visual properties. Users are interested in the images placed at the first positions of the returned ranked lists, which usually are the most relevant ones. Therefore, accurately ranking collection images is of great relevance. However, in general, CBIR approaches perform only pairwise image analysis, that is, they compute similarity (or distance) measures considering only pairs of images, ignoring the rich information encoded in the relationships among images. Aiming at improving the effectiveness of CBIR systems, re-ranking and rank aggregation algorithms have been proposed. Re-ranking algorithms have been used to exploit contextual information, encoded in the relationships among collection images, while rank aggregation approaches have been used to combine results produced by different image descriptors. In the work developed by the principal investigator during his PhD research, several methods have been proposed for image re-ranking and rank aggregation, aiming at improving the effectiveness of CBIR systems. Experimental results demonstrated the effectiveness of the proposed approaches in comparison with other state-of-the-art methods recently proposed in the literature. However, the relevant results obtained led to new important research challenges. The objective of this research project is to investigate the re-ranking and rank aggregation approaches under various aspects, addressing the challenges still open. Important aspects to be investigated are related to the scalability and efficient computation of image re-ranking algorithms using parallel algorithms on heterogeneous computing environments. Another relevant aspect is the specification and implementation of new re-ranking approaches to be used in different scenarios and applications, such as multimodal and textual retrieval, relevance feedback, and collaborative image retrieval.

---

Large-Scale Active Learning for Visual Retrieval.
Royal Academy of Engineering - Newton Research Collabration Programme

2015-2017
Principal Investigator: Ying Weng (Bangor University - UK)

Abstract:
A massive and ever growing amount of digital image and video content is available today, on websites such as YouTube and Flickr, in archives such as BBC and NBC, and in personal collections. In fact, a change of behaviour can be observed, since common users are not long mere consumers and have become active producers of digital multimedia content. In most cases, it comes with additional information, such as text or other metadata, that forms a rather sparse and noisy, yet rich and diverse source of annotation. Nevertheless, there are various challenges involved in the retrieval tasks. While the text-based retrieval models are well established, they ignore the rich source of information encoded in the visual and audio data. On the other hand, promising content-based retrieval technologies, although capable of considering the multimedia content, still face obstacles for mapping the low level features into high level semantic concepts. Supervised approaches, as relevance feedback techniques for example, have been employed for visual and multimodal retrieval. Although very effective, such methods require a lot of user intervention. In this scenario, the retrieval approaches are ideally suited to emerging weakly supervised and active machine learning technology. Therefore, this project aims at autonomously exploring data collections, employing unsupervised learning techniques for considering the relationships among multimedia objects and saving the user's efforts. Considering that parallelization is becoming necessary due to multicore processors and the high availability of GPGPUs with thousands of threads, this project also aims at applying parallelization strategies to improve the efficiency of the proposed methods.