JosephKJ/OWOD

Problems about ablation experiments and Oracle

moyoung1996 opened this issue · 1 comments

moyoung1996 commented

Thanks for your great paper! I have some questions about the paper.
1.What does Oracle actually mean? Is it an existing detector or something else? I cannot find any paper about it.
2.Row 4 of ablation experiments(Table 4) shows that only ALU can have a good effect on experiments. How it works without contrastive clustering?

JosephKJ commented

Hi @moyoung1996, thank you for your kind words.

  1. What does Oracle actually mean? Is it an existing detector or something else? I cannot find any paper about it.
  • As mentioned in Sec 5.3: "An ‘Oracle’ detector has access to all known and unknown labels at any point, and serves as a reference."
  1. Row 4 of ablation experiments(Table 4) shows that only ALU can have a good effect on experiments. How it works without contrastive clustering?
  • The RPN proposals are auto-labelled based on the objectness measure. Please find more info in Sec 4.2. Contrastive clustering is not a pre-req for ALU.