Images extracted from the datasets used to train the method.
Video understanding is the next frontier of computer vision, in which activity recognition plays a major role. Ph.D. student Victor Hugo Melo will present his new approach to the subject, based on contextual cues obtained from object detections in the scene, at the main conference on computer vision in South America, the Conference on Graphics, Patterns and Images, SIBGRAPI 2018. In its 31st edition, the conference will be held from October 29th to November 1st in Foz do Iguaçu, Paraná.
In his work, Melo leverages contextual cues obtained from object detections in the scene, assuming that objects relevant to an action present characteristic spatial relationships with respect to the subject performing it. When combined with a state-of-the-art convolutional neural network, the results are promising, showing that the proposed method improves the recognition of certain object-oriented actions. “Our experiments support the hypothesis that the spatial disposition of an object provides an important clue to activity recognition” says the researcher.
The data-driven approach that will be presented at the conference shows that besides such spatial features, there may be other important information that further enhances object-based activity recognition, such as co-occurrence, relative size, and temporal information.
About this participation, Victor Hugo stated, “presenting this work at SIBGRAPI 2018 is a great opportunity, as it will allow us to get feedback from the community, looking for novel ways on how to approach the problem and to extend this work”.
Melo holds a Master’s degree in Computer Science from UFMG with the dissertation “Fast and Robust Optimization Approaches for Pedestrian Detection”. His main research topics are computer vision and human activity recognition. He is also a technical leader in R&D projects with private companies.
Led by Melo, the work counted with contributions from Smart Sense Laboratory undergraduate student Jesimon Barreto dos Santos, graduate students Carlos Caetano and Jessica Sena, Otavio Penatti from Samsung Research Institute and Professor William Schwartz, Melo’s advisor and coordinator of the Smart Sense Laboratory.
The Conference on Graphics, Patterns and Images is a leading annual event combining contributions from four major subjects related to image computing: computer graphics & vision, pattern recognition and image processing. The proceedings have been published by IEEE Xplore since 1997. This year SIBGRAPI will be co-located with two other major events: SBGames 2018, the Symposium on Computer Games and Digital Entertainment, and SVR 2018, the Symposium on Virtual and Augmented Reality.