Evaluation on UnRel

[nickgkan]

Hi, thanks for the code!

I cannot figure out how to evaluate retrieval on UnRel. Assuming I have trained a relationship detector on VRD. I run inference on both VRD and UnRel images, which gives me scores for 70+1 classes for every pair of object. How do we compute the retrieval mAP now? Should I keep a single score per pair?

I think I'm missing some serious points here, could you explain?