I've been recently using the YOLO system to detect people on images, which turned out really well.
My next step is to try to find images of the same person across the whole set of images I've retrieved earlier. I came across the AlignedReID: Surpassing Human-Level Performance in Person Re-Identification paper, read it through and found an implementation of it that offers pre-trained weights.
However, I'm kind of lost when trying to apply it to my (unlabelled) custom dataset. To me, the input of the system is an image and its outputs are k candidate images, ranked accordingly to the system's confidences. In other words, the system will try to match an image with the other images he works with.
Since the images I retrieved are unlabelled, I'd like to avoid re-training anything. But the more I think of it, the less it seems possible to me.
Can I load the pre-trained weights, define the images I retrieved as the test set and run it? Or should I re-train the last layers to adapt it to my dataset?
In other words, how generalisable are person re-identification systems?