Computer Vision News - May 2023

22 Women in Computer Vision some classes. Or, I think that it's from this geographic region. The hierarchical notion of uncertainty is something that I've really been thinking a lot about lately. It's not super well integrated into a lot of the image retrieval approaches that exist. But I think it would make our tool much more powerful for investigators rather than just always returning an image search result. Here's a similar image that's not actually that similar. If we could say I don't actually know what a similar image is, but I sure think that that is from New York City. Then we would explain to an investigator why. We have work that I've published at WACV, going back a few years now, trying to explain image similarity results. I get two images that I say are similar. How can I explain to an investigator why the model anthropomorphizes. I think we've made heat maps as the approach so far. It says about the system that you have put in place. Is there one thing that you would like your computer to tell you, and it does not tell you yet? [ laughs ] That's a great question. I would love to have much better models of uncertainty. This is something we've been thinking about in recent work that is not published yet but hopefully will be soon. In our image search tool for investigators, the way we present results is we say you gave us this image. Here are the thousand images that we may think are most similar to the hotels that they came from, but that have no kind of notion of uncertainty. And I think it would be a much more useful system for investigators if we could say, I actually don't know what image is most similar, but I really think it's this particular hotel chain, right? I don't have a good image retrieval match, but I have a reasonable estimate of