RelTR code reading

https://github.com/yrcong/RelTR https://github.com/long8v/PTIR/issues/40

  • Understand how bipartite matching works
  • The final output will look like (batch_size, num_of_triplets, -1).
  • If we define a cost function (=classification error with gt + bbox error) in each dimension (bbox, since the classification output will be separate) and summarize it, it will look like (batch_size, num_of_triplets, cost) and use the scipy package to get an index that minimizes cost and matches all gt!
  • What happens to inference then??? -> this is DETR again