CVPR Daily - 2018 - Thursday

Guha Balakrishnan, a postdoc at MIT; Adrian Dalca, a postdoc between MIT and Harvard; and Amy Zhao, a PhD candidate at MIT, speak to us about their work ahead of their oral and poster session today. Guha tells us that the work is about taking an image of a person and synthesising a new realistic image of that person in a different pose . In the real world, this could be a useful way of augmenting datasets that already exist and as a component of learning systems that need to imagine the world in different scenarios. Adrian says that it can also be used as a way to teach people how to do something. For example, to improve their golf swing by synthesising it. Guha says that one of the challenges of this type of problem is that in order to synthesise a realistic image, you have to take into account quite large changes in the image due to movement, as well as complex appearances in the background and the foreground when you work with natural images. This method is specifically aimed to face those challenges. Guha explains what kind of algorithmic techniques they have used: “ We take a learning-based approach. The main contribution of our approach as opposed to similar related work is that our learning method is very modular. We split the problem into four different smaller subtasks that are intuitively how we 10 Guha, Amy, Adrian Thursday From left: Adrian Dalca, Guha Balakrishnan andAmy Zhao