Realization of SPRC2

[zyn99xjtu]

I am particularly intrigued by the SPRC2 model described therein. While I have grasped the fundamental architecture and tasks of the model, I found some challenges in reproducing the results using SPRC2. Therefore, I want to request more detailed information specifically on the implementation of the re-rank method of the SPRC2. Does it use the extra model same as "Candidate set re-ranking for composed image retrieval with dual multi-modal encoder.", or change to a Q-Former-like structure?

I would be immensely grateful if you could provide SPRC2's code or some detailed steps on how to reproduce the performance on CIRR, and I also wonder why there is no result on FashionIQ using SPRC2.

Thank you so much.