Learning similarities between images is a fundamental problem in Computer Vision and Image Analysis. I am interested in framing the problem of similarity learning within the Convolutional Neural Networks framework in order to learn transitivity properties over image similarities.
Human Pose Recovery using contextual representations
A major challenge for true AI to become a reality is to be able to interpret and understand humans and their interactions. In this sense, estimating the human pose in RGB images is a crucial task for further development in the field.
Recent breakthroughs in Convolutional Neural Networks have boosted the Computer Vision field enormously. However, like most supervised classification methods, they assume categories live in a equidistant label space. I am interested in studying which configurations of the label space are more suitable to further boost CNNs expressive properties.
Drivable surface detection
Autonomous driving is currently an extremely hot topic. However, autonomous driving systems (Tesla, Google, Uber, etc.) are based in very expensive sensor arrays (Lidar, GPS, laser arrays, etc.). I am interested in using unexpensive RGB images to model which areas of an image are drivable.
Computer Vision for Art-historical images (funded by Heidelberg Univerity - WinKolleg)
In this open we open a dialogue about artistic recognition and similarity, as well as, case studies in a wide range of artistic productions which combines the two fields in the context of Computational Humanities