Are you aware of the sub-graph set of methods?

[wishforgood]

There are some papers using sub-graph proposal instead of relation proposal. However, they perform experiments on other subset of Visual Genome, and is hard to compared with the methods mentioned in this repo. Here are the links of papers I found. http://openaccess.thecvf.com/content_ECCV_2018/papers/Yikang_LI_Factorizable_Net_An_ECCV_2018_paper.pdf https://arxiv.org/pdf/1904.00560.pdf

[jwyang]

Hi, @wishforgood , thanks for reminding, Yes I knew these papers. but as you mentioned, it is indeed hard to compare, since the data splits are different.