After reading the DFC guide, I wasn’t able to find a clear explanation as to what exactly the calibration dataset is supposed to be. I have a couple questions below.
Is it supposed to contain labels, or just the images? If it needs labels, what format?
Should the calibration images be from the training set or should they just be images that best resemble the target inference images?
Hi @esahin,
No labels are needed, the calibration is a way to collect statistics on the data that the network is going to process, thus allowing for a smarter choise of the quanlized bins. The statistics are collected per layer.
Both would probably be good. but it should be represantative, so if you have a real video from a scene, taking 100 consective images would not be a good choice.