Computer Vision News - July 2019

you devise an unsupervised loss which is just an entropy loss to update the scale and the bias for the domain- specific parameters. Massimiliano tells us that they were restricted in terms of the data they had to test on. They tested on the data set used in the literature for this problem, CompCars, with different cars, with a different viewpoint of cars year of production. Also, another data set which depicted different portraits collected over 100 years in different regions of America. This is not a real application of this problem. He says they would like to have a model, for example, in autonomous driving. He explains: “ I tell you it’s evening, it starts getting darker, so it adapts your model for this kind of light. Obviously, one can say that if I have a huge amount of data which are balanced among all the possible conditions. Our algorithm does not do it. Nowadays, the data sets are unbalanced, and we must specialize our systems. ” Massimiliano says they assume the metadata is representative of the domain shift, which cannot be true, so they must weight them. He says a next step for this work would be to understand which metadata are important for the domain shift and which are not. He also thinks they could go from the metadata to a more abstract representation, because metadata is good if you have it and if you can quantify the shift, but if someone tells you it’s dark or it’s darker, you can understand what the environment will look like, so a description rather than metadata would be helpful. Finally, Massimiliano is excited to tell us that he is very proud of his heritage: “ I come from a very small village in the middle of Italy in Umbria. I think we are 30 people in the centre of the village. It was a long journey to get here and I’m very happy that at the end, even starting from there, I can represent my small village here .” 18 DAILY CVPR Wednesday Presentation " Our graph is able to estimate target model parameters with pretty good results in our experiments. Receiving the target data and using them to refine our model, allows to fill the remaining gap with the upper bound, a standard domain adaptation algorithm with target data available beforehand."