When you snap a picture outside, the weather can be all kinds of things - sunny, rainy, cloudy, you name it. Weather really affects how well computer programs can figure out what's in a photo. Most current methods treat weather as a clear-cut thing, like a list of choices. But in the real world, weather isn't so simple. It's always changing, and it can even mix up - like when it's both sunny and cloudy at the same time.

Instead of just picking one weather type, we should look at how weather is really like a blend of things. That's why we're working on a new way to handle weather in computer vision tasks. We're starting by rethinking what weather really is, based on the rules of physics. We're calling this "modeling weather uncertainty. "

To do this, we're using something called a Gaussian mixture model. It helps us see the blend of different weather types in a photo. And we're also coming up with a new way for computers to learn, using a concept called "prior-posterior learning. " This is like how a detective would tell you the chances of something happening after seeing all the clues.

We're also making a new dataset to test out our method. It has 14 different weather categories and over 16, 000 photos. Plus, we're using a new type of computer brain called a "MeFormer" to help figure out the weather blend in a photo.

Our tests show that our method works really well on both the usual weather guessing tasks and our new blend-spotting task. Plus, it can even help computers figure out what's in a photo when the weather is really bad.

Weather's Continuous Shift: A New Way to See It

Actions