Leo Sampaio Ferraz Ribeiro

• Abstract

  The spread of Child Sexual Abuse Material (CSAM) over the internet is an ever-growing, worldwide problem that re-victimizes its victims and becomes increasingly intractable with current manual checking and investigative efforts from law enforcement agencies. Furthermore, the burden of dealing with this kind of data wears heavily on agents that have to categorize hundreds of terabytes of d…

  Few-Shot Scene Recognition, a New Proxy Task for Child Sexual Abuse Material Detection, BP.PD

• Abstract

  Feature learning methods have reached state of the art on many areas of study. Despite excellent results achieved on benchmark datasets, there is still little understanding of its inner workings, and applications to be explored, particularly when considering architectures that go beyond traditional convolutional neural networks. In this project, we propose the use of feature learning for …

  Generative networks and feature learning for cross domain visual search, BP.DD

• Abstract

  Siamese and triplet-loss-based nets, both variations of Convolutional Neural Networks (CNNs), are commonly applied to Content-based Image Retrieval and Domain Mapping systems thanks to their natural characteristic of generating feature spaces tuned for retrieval. There is however an improvement opportunity lying on the representation of the generated data; focusing on limited resources pl…

  Dimensionality reduction methods for representations generated by triplet convolutional networks, BP.IC

(Only some records are available in English at this moment)

• Abstract

  Current advances within cross-domain image retrieval are mostly dependent on the availability of large amounts of annotated data. In this project we seek to explore the use and development of unsupervised feature learning methods that can overcome the natural domain gap when dealing with images from different styles and domains. We are particularly interested in representations infered by…

  Finding correspondences between cross-domain representations using unsupervised feature matching, BE.EP.DD

• Abstract

  There are image classification problems in which each image is represented by regions of interest; for each region a series of features can be extracted. As a result a set of feature vectors are available, and it is necessary to assign a label to this set of instances. The Multiple-Instance Learning (MIL) studies the problem in which each object is described as a bag (set of instances). T…

  Multiple-instance image ranking for sketch-based image retrieval, BE.EP.IC