CVPR Daily - Wednesday

Wednesday 13 Emanuela Marasco specifically, but I’m interested in applying this research on additional modalities as well which are iris and face. Regarding the cross sensor matching, a few years ago, I think in 2015, I started a research through which I found out, even in the same technology, fingerprint sensors certified by the FBI, the ones that are used at the airport, for example, for US visas, are actually biased by the sensor: the matching between the fingerprint required by different devices is different. There is an increase of error rate in cross sensor scenarios. What we are doing is developing machine learning-based approaches to reduce this impact on the performance. I tried to understand the variations due to device diversity in these images, but there is no model for that. So now I am working my interests in developing models. We are applying autoencoders in order to generate a robust representation of the images that can be matched correctly, even across different sensors. Between iris, face, and fingerprint, which are the most fascinating to work with? Well, I’m still very fascinated by fingerprint and iris as well. I’m told that I have almost no fingerprint. Is that possible? Well, yes. There are physical factors that your fingerprint can not be read very well from the sensor. I think this is the question. The configuration is stable so it doesn’t change. Then the way you interact with the touch-based sensor, that is the difference in matching. The quality depends on the minutiae point. Minutiae points are the points that we extract from the configuration of the fingerprint. We have a minutiae for the point, the location where the ridges (the black lines) bifurcate, and where they end. There is also indentation information involved. With aging, the skin becomes dryer so the contact is not very good. [ laughs ] I expected a better reason for why I don’t have fingerprints! [ laughs too ] Also having a manual job can impact the condition of the finger over time. Also, if they get wet, the image becomes darker and you cannot recognize the minutiae point in the image. One other aspect that I like about fingerprint is, of course, the gender estimation. Mainly, the configuration is randomly determined, but I did read many medical reports