Computer Vision News - July 2019

the edges, where the weight of the edge basically says how much two domains are related. If I tell you I see a car from a side view and from a rear- side view, a rear side and side are closer than a rear and front, so we relate in this way. When we get the target attribute, at that point we initialize a virtual node – a fake node – because we didn’t have any data for that. No parameters are there, but if you assume that similar metadata, so similar domains in our graph – similar nodes – require similar parameters, we can just propagate the parameters of nearby nodes and obtain our model for the target domain which we never see .” He adds that there is another problem with that. Like before, if someone tells you that it’s going to rain, you go out with an umbrella, but if they don’t tell you that and you go out and it starts raining, then you must figure out how to react. Since obviously there will be nothing supervised, and so the prediction of the target model can be wrong of course – and also, some metadata may be received which are not representative of the domain – this method unifies this kind of prediction with continuous domain adaptation. As the target data is received, the model is continuously updated. This is possible because the model is based on batch normalization. The different domains have different batch normalization statistics and scale and bias parameters. For each domain, there are different statistics, and so for the target you basically predict just the statistics of the domains and the scale and the bias. Once you have those, at test time when you receive the target data, you use it to update the statistics because it can be done easily. Then 17 DAILY CVPR Wednesday Massimiliano Mancini " This is possible because the model is based on batch normalization "