Implementing a new dataset to CV architecture


I want to implement the PIE dataset in the AgentFormer arch.
AgentFormer uses ETH and nuScene datasets. I successfully run these datasets on this arch. However, I couldn’t take a good way with the PIE dataset. I am not sure how I could write a new data loader for it. Which steps should I take?
I normally have experience in implementing articles without looking at similar GitHub codes, but this time I stuck.

Thank you for any help.