Original article was published on Artificial Intelligence on Medium
Unifying visual embeddings for visual search at Pinterest
Their visual cropping tool, as well as search, is pretty amazing, I used it and it works really well.
And it seems like how they are able to achieve this is via good embedding. (another form of finding a good manifold).
There is a different set of the dataset and each one of them has a specific use case. Very interesting!
They want to unify the visual embedding, for both 3D and 2D and more.
The thing is how can we select images that are negative? What does that even mean?
And more importantly why is BAG negative to a person wearing a BAG?
They way they are trying to solve this problem is by designing a new loss function. Multiple loss functions, this follows the general paradigm in deep learning, in my opinion. (finding a better manifold/create a creative loss function).