search

Jingkang50 / OpenPSG

Benchmarking Panoptic Scene Graph Generation (PSG), ECCV'22
https://psgdataset.org
MIT License
407 stars 68 forks source link

About table result #89

Closed sirosann closed 1 year ago

sirosann commented 1 year ago

Hello I have a issues about a description in the paper. After table 2 writed [ For example, even the most classic method IMP can reach over 30% R@20, which already exceeds all the available R@20 under the scene graph generation (SGDet) task (cf . 28.4% by PSGTR). ] I wonder why the IMP predicate classification results are compared to PSGTR SSG results here.

thank you

Jingkang50 commented 1 year ago

Thank you for the question. Compared to SGDet, SGCls is simpler since GT masks are provided. (IMP's SGDet << PSGTR's SGDet) but (IMP's SGCls > PSGTR SGDet) show the accurate masks in the first stage might be important to get a high final SGDet score.

sirosann commented 1 year ago

Thank you for your answer. I got it .