Descriptors with float values

[krsapkota]

I am using a gpu implementation of SURF to get descriptors and it gives me descriptor values in float [-1 1]. If I scale that to [0 255] I lose a lot of precision. Is there a way I can train the Vocabulary tree on float descriptor values instead of integer values?

Thanks a lot!

[snavely]

I don't currently have plans to add support for floating point descriptors (but feel free to try! Beyond changing the datatypes, I don't think this would be hugely complicated. Just some refactoring.) That said, even if you mapped your features from [-1, 1] to [0, 255] and quantized, I bet it would work pretty well despite the loss of precision.