Computer Vision News - July 2019

Massimiliano Mancini is a PhD student at Sapienza University of Rome, in collaboration with Fondazione Bruno Kessler (FBK) and Istituto Italiano di Tecnologia (IIT). His advisors are Barbara Caputo from IIT and Politecnico of Torino, and Elisa Ricci from the University of Trento and FBK. Samuel Rota Bulò from Mapillary is a third unofficial advisor. Massimiliano spoke to us ahead of his oral and poster. This work tackles predictive domain adaptation. In standard domain adaptation, what you have is a source domain where you have a lot of labels for the task that you want to address, then another domain, which is unlabelled, which is actually the domain which you want to apply your model. For instance, you have a lot of images collected in daylight, so they are clean and labelled, and then you want to apply your model at night, but you have only unlabelled images for the night situation. You must figure out how to go from daylight to night without any label on the night. You need the target data for doing this task because if you don’t see anything you cannot forecast what night will look like. This work tries to predict what the night, or a situation, will look like. If you are given a lot of domains – one is labelled, a lot are not labelled – for each of them, you are given an attribute. For example, you see a front view of a car, and then you see from the side, and then someone asks you to produce a model that will recognise cars if you see them from the back. Since you are seeing a lot of different domains, either labelled or unlabelled, you may try to figure out what the model for the rear of the car will look like. This is something that we do even as humans. If someone tells you that it will rain tonight, you might go out with an umbrella because you adapt yourself to the weather conditions. If they don’t tell you it’s going to rain, you might go out with nothing. As humans, we adapt to what we know, and if we have understood how the weather changes with respect to some attributes, we can try to adapt to that. We hope that our algorithms are able to do that too. Massimiliano explains: “ To solve this task you must relate what the parameters of a domain are with respect to the attribute, or the metadata, in our case. We have a first phase where we train parameters which are attribute or domain specific, so they are specific for a certain condition. While doing that, we initialise what we call a graph. We have a node for each of the domains and we connect the nodes of the graph with AdaGraph: Unifying Predictive and Continuous Domain Adaptation Through Graphs 16 DAILY CVPR Wednesday Presentation