O professor da Université Bourgogne Europe in Dijon, France, Renato Martins, irá ministrar, dia 28/08, sexta-feira, às 10 horas, na sala 2077 do Instituto de Ciências Exatas (ICEx) da UFMG, a palestra “Improving Visual Feature Learning with Geometric and Semantic Priors”. A palestra faz parte dos Seminários Avançados da Pós-graduação, mas é aberta ao público, sem necessidade de inscrição prévia e gratuita. Abaixo o resumo da palestra e bio do professor:
Titulo: Improving Visual Feature Learning with Geometric and Semantic Priors
Resumo: Geometric and semantic cues can provide valuable information for guiding the learning of features that are better adapted to the underlying structure of visual data. In this talk, I will discuss how such priors can be leveraged to improve visual feature learning with two recent results. First, I will present our recent work on omnidirectional video interpolation (SVI360), where equivariant geometric constraints are incorporated to synthesize intermediate frames of large field-of-view images. I will then present how clustering visual features with prototypes within 3D Gaussian Splatting representations can improve camera pose regression regarding scene scale and illumination changes.
Bio: Renato Martins is an Assistant Professor (Maître de Conférences) at Université Bourgogne Europe in Dijon, France. He received his Ph.D. in Computer Science at INRIA Sophia Antipolis and Mines Paris – PSL, under the supervision of Dr. Patrick Rives. He subsequently held postdoctoral positions at INRIA Sophia Antipolis and in the Department of Computer Science at the Universidade Federal de Minas Gerais (Brazil), where he worked on 3D computer vision with Prof. Erickson R. Nascimento. His main research interests include computer vision and machine learning for visual perception, with a particular focus on visual matching, human motion analysis, and 3D reconstruction of rigid and deformable scenes. For more details, please visit his personal webpage: https://renatojmsdh.github.io/
