Papers by Cristina Vasconcelos
2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Figure 1. Impact on object detection performance of preserving and not fine-tuning features learn... more Figure 1. Impact on object detection performance of preserving and not fine-tuning features learned on ImageNet. The arrows indicate impact on each model, when trained with frozen backbone weights. As long as the remaining detector components have enough capacity (left), freezing increases performance while significantly reducing resources used during training (right).
Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 2018
Games and other applications are exploring many different modes of interaction in order to create... more Games and other applications are exploring many different modes of interaction in order to create intuitive interfaces, such as touch screens, motion controllers, recognition of gesture or body movements among many others. In that direction, human motion is being captured by different sensors, such as accelerometers, gyroscopes, heat sensors and cameras. However, there is still room for investigation the analysis of motion data captured from low-cost sensors. This article explores the extent to which a full body motion classification can be achieved by observing only sparse data captured by two separate inherent wereable measurement unit (IMU) sensors. For that, we developed a novel Recurrent Neural Network topology based on Long Short-Term Memory cells (LSTMs) that are able to classify motions sequences of different sizes. Using cross-validation tests, our model achieves an overall accuracy of 96% which is quite significant considering that the raw data used was obtained using only 2 simple and accessible IMU sensors capturing arms movements. We also built and made public a motion database constructed by capturing sparse data from 11 actors performing five different actions. For comparison with existent methods, other deep learning approaches for sequence evaluation (more specifically, based on convolutional neural networks), were adapted to our problem and evaluated.
ArXiv, 2021
Adversarial robustness is an open challenge in deep learning, most often tackled using adversaria... more Adversarial robustness is an open challenge in deep learning, most often tackled using adversarial training. Adversarial training is computationally costly, involving alternated optimization with a trade-off between standard generalization and adversarial robustness. We explore training robust models without adversarial training by revisiting a known result linking maximally robust classifiers and minimum norm solutions, and combining it with recent results on the implicit bias of optimizers. First, we show that, under certain conditions, it is possible to achieve both perfect standard accuracy and a certain degree of robustness without a trade-off, simply by training an overparameterized model using the implicit bias of the optimization. In that regime, there is a direct relationship between the type of the optimizer and the attack to which the model is robust. Second, we investigate the role of the architecture in designing robust models. In particular, we characterize the robustn...
ArXiv, 2020
Image pre-processing in the frequency domain has traditionally played a vital role in computer vi... more Image pre-processing in the frequency domain has traditionally played a vital role in computer vision and was even part of the standard pipeline in the early days of deep learning. However, with the advent of large datasets, many practitioners concluded that this was unnecessary due to the belief that these priors can be learned from the data itself. Frequency aliasing is a phenomenon that may occur when sub-sampling any signal, such as an image or feature map, causing distortion in the sub-sampled output. We show that we can mitigate this effect by placing non-trainable blur filters and using smooth activation functions at key locations, particularly where networks lack the capacity to learn them. These simple architectural changes lead to substantial improvements in out-of-distribution generalization on both image classification under natural corruptions on ImageNet-C [10] and few-shot learning on Meta-Dataset [17], without introducing additional trainable parameters and using the...
Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 2018
Mixed reality is the union of virtual and real elements in a single scene. In this composition, o... more Mixed reality is the union of virtual and real elements in a single scene. In this composition, of real and virtual elements, perceptual discrepancies in the illumination of objects may occur. We call these discrepancies the illumination mismatch problem. Recovering the lighting information from a real scene is a difficult task. Usually, such task requires prior knowledge of the scene, such as the scene geometry and special measuring equipment. We present a deep learning based technique that estimates point light source position from a single color image. The estimated light source position is used to create a composite image containing both the real and virtual environments. The proposed technique allows the final composite image to have consistent illumination between the real and virtual worlds, effectively reducing the effects of the illumination mismatch in Mixed Reality applications.
Entertainment Computing and Serious Games, 2019
Lecture Notes in Computer Science, 2018
Deep learning models show remarkable results in automated skin lesion analysis. However, these mo... more Deep learning models show remarkable results in automated skin lesion analysis. However, these models demand considerable amounts of data, while the availability of annotated skin lesion images is often limited. Data augmentation can expand the training dataset by transforming input images. In this work, we investigate the impact of 13 data augmentation scenarios for melanoma classification trained on three CNNs (Inception-v4, ResNet, and DenseNet). Scenarios include traditional color and geometric transforms, and more unusual augmentations such as elastic transforms, random erasing and a novel augmentation that mixes different lesions. We also explore the use of data augmentation at test-time and the impact of data augmentation on various dataset sizes. Our results confirm the importance of data augmentation in both training and testing and show that it can lead to more performance gains than obtaining new images. The best scenario results in an AUC of 0.882 for melanoma classification without using external data, outperforming the top-ranked submission (0.874) for the ISIC Challenge 2017, which was trained with additional data.
Revista de Informática Teórica e Aplicada, 2011
A Unidade de Processamento Gráfico-do inglês "Graphics Processing Unit"(GPU) foi desenvolvida ini... more A Unidade de Processamento Gráfico-do inglês "Graphics Processing Unit"(GPU) foi desenvolvida inicialmente como um hardware destinado a aumentar a eficiência e o poder de processamento gráfico para tarefas de renderização. Hoje, a GPU apresenta-se como um hardware de processamento versátil e de alto poder de computação. Tornou-se uma possibilidade real na busca por soluções para processamento em grandes volumes de dados, seja como complemento, seja como alternativa ao uso de CPUs multicore ou de sistemas distribuídos. A utilização da GPU em computações de propósito geral é de especial interesse, uma vez que para diversas aplicações, ainda não existem formulações sequenciais suficientemente rápidas de serem computadas. Este tutorial tem como objetivo permitir ao leitor a identificação de algoritmos e aplicações candidatas à abordagens paralelas em GPU. Com tal finalidade, apresentamos os fundamentos e conceitos envolvidos na programação de propósito genérico utilizando hardware gráfico sem que seja indispensável ao leitor, o conhecimento a priori de sistemas gráficos 3D ou de sistemas paralelos.
Unsupervised cosegmentation based on global clustering and saliency
2015 IEEE International Conference on Image Processing (ICIP), 2015
This paper introduces a new method for unsupervised cosegmentation. Our method combines saliency ... more This paper introduces a new method for unsupervised cosegmentation. Our method combines saliency information with a Global Clustering step, which reveals parts of the objects by detecting similar subregions across image collections, based on a low dimensional descriptor that includes color, texture and positional features. The saliency information is used to yield a classification of the global clusters into foreground and background and also classify regions not detected as global clusters into potential background or foreground. These four types of regions are the input seeds for a Graph Cuts procedure that computes the final cosegmentation. The Graph Cuts result can also be used to compute a refined version of the saliency information which enables us to define an iterative cosegmentation pipeline. Our framework produces remarkable results in comparison with state-of-the-art works, even in challenging datasets with illumination variance, occluded objects and identical background.
Observed Interaction in Games for Down Syndrome Children
2015 48th Hawaii International Conference on System Sciences, 2015
This work proposes a method for evaluating the children's behavioral interactions with a game... more This work proposes a method for evaluating the children's behavioral interactions with a game, more specifically for evaluating playful applications for kids with cognitive disabilities. Our method introduces an evaluation criteria over children's behavioral interaction and game design analysis, adapted from a list of breakdown indication types of the Detailed Video Analysis (DEVAN) that was originally designed for regular applications. We present a case study of the proposed evaluation method with a detailed analysis of the game called JECRIPE, originally developed for stimulating cognitive abilities of children with Down syndrome in preschool age. The proposed method adopts qualitative and quantitative criteria to review the initial developmental factors that have driven JECRIPE's design versus the real behavior observed in a group of children playing the game. As results of this case study, we demonstrate the reliability of the evaluation method and the capacity of this method in discovering usability and fun problems in order to be considered and addressed in future game releases.
Evaluating and Customizing User Interaction in an Adaptive Game Controller
Lecture Notes in Computer Science, 2015
ABSTRACT
Multi-thread architectures are the current trends for both PCs (multi-core CPUs and GPUs) and gam... more Multi-thread architectures are the current trends for both PCs (multi-core CPUs and GPUs) and game consoles such as the Microsoft Xbox 360 and Sony Playstation 3. GPUs (Graphics Processing Units) have evolved into extremely powerful and flexible processors, allowing its use for processing different data. This advantage can be used in game development to optimize the game loop. As reported in the literature, GPGPUs have been used in processing some steps of the game loop, while most of the game logic is still processed by the CPU. This proposal differs by presenting an architecture designed to process practically the entire game loop using the GPU. Two test cases, a crowd simulation and a 2D game shooter prototype called GpuWars, are presented to illustrate the proposed architecture.
This paper presents MOCT, a multi-object chromatic tracking technique for real-time natural video... more This paper presents MOCT, a multi-object chromatic tracking technique for real-time natural video processing. Its main step is the MOCT localization algorithm, that performs local data evaluations in order to apply a multiple output parallel reduction operator to the image. The reduction operator is used to localize the positions of the object centroids, to compute the number of pixels occupied by an object and its bounding boxes, and to update object trajectories in image space. The operator is analyzed using three different computation layouts and tested over several reduction factors.
Jecripe
Proceedings of the 7th International Conference on Advances in Computer Entertainment Technology, 2010
Digital games are usually developed to provide fun for people of all ages. Although games have be... more Digital games are usually developed to provide fun for people of all ages. Although games have been mostly used for entertainment purposes, they have great potential as an intervention tool in health care. Digital games can be applied in health care helping users to learn or to experience something in a fun way. However, there are important issues to be considered to achieve this goal, specially in the development of applications for people with special needs. In this work, we are concerned specifically with Down syndrome needs, and we present an unhackneyed game for children with Down syndrome between 3 to 7 years old. Children in pre-scholar age need to be stimulated considering different cognitive areas. The stimulation of such cognitive areas can provide good results in the development over the years. Given the current demand and absence of games that fulfill Down syndrome special needs, we developed JECRIPE. JECRIPE is a digital game that stimulates the specific cognitive abilities: imitation, perception, fine motor skills, hand-eye coordination and receptive and expressive verbal language. This paper describes how these cognitive areas are used to stimulate children with Down syndrome in pre-scholar age and some technical issues in the development of the game.
2010 Brazilian Symposium on Games and Digital Entertainment, 2010
Figure 1: JECRIPE-a game for children with special needs Abtract There are not many initiatives i... more Figure 1: JECRIPE-a game for children with special needs Abtract There are not many initiatives in the area of game development for children with special needs, specially children with Down syndrome. The major purpose of our research is to promote cognitive development of disabled children in the context of inclusive education. In order to do so, we address aspects of interaction, communication and game design in stimulating selected cognitive abilities. By using a Human-Computer Interaction method based on the Inspection of Evaluation, it was possible to study and understand user interaction with the interface and thus examine the positive aspects as well as the communicability problems found with the JECRIPE game-a game developed specially for children with Down syndrome in pre-scholar age.
Bipartite Graph Matching on GPU over Complete or Local Grid Neighborhoods
Lecture Notes in Computer Science, 2011
ABSTRACT
Lecture Notes in Computer Science, 2009
The Bipartite Graph Matching Problem is a well studied topic in Graph Theory. Such matching relat... more The Bipartite Graph Matching Problem is a well studied topic in Graph Theory. Such matching relates pairs of nodes from two distinct sets by selecting a subset of the graph edges connecting them. Each edge selected has no common node as its end points to any other edge within the subset. When the considered graph has huge sets of nodes and edges the sequential approaches are impractical, specially for applications demanding fast results. In this paper we investigate how to compute such matching on Graphics Processing Units (GPUs) motivated by its increasing processing power made available with decreasing costs. We present a new data-parallel approach for computing bipartite graph matching that is efficiently computed on today's graphics hardware and apply it to solve the correspondence between 3D samples taken over a time interval.
2013 XXVI Conference on Graphics, Patterns and Images, 2013
Fig. 1: Example of the stages of our method applied to a pixel art based game. The first image sh... more Fig. 1: Example of the stages of our method applied to a pixel art based game. The first image shows the input image scaled 16x with nearest neighbor. The second image represents the similarity graph for the image. The third image shows how pixels are reshaped into cells and the last image is the final result with the smoother triangulated cells (Original image c Sega Corporation).
Lloyd’s Algorithm on GPU
Lecture Notes in Computer Science, 2008
Page 1. Lloyd's Algorithm on GPU Cristina N. Vasconcelos1, Asla Sá2, Paulo Cezar... more Page 1. Lloyd's Algorithm on GPU Cristina N. Vasconcelos1, Asla Sá2, Paulo Cezar Carvalho3, and Marcelo Gattass1,2 1 Depto. de Informática - Pontifıcia Universidade Católica (PUC-Rio) 2 Tecgraf (PUC-Rio) 3 Instituto de ...
Uploads
Papers by Cristina Vasconcelos