jwyang / graph-rcnn.pytorch

[ECCV 2018] Official code for "Graph R-CNN for Scene Graph Generation"
733 stars 157 forks source link

Question of Data size #72

Closed aLMonotonous closed 4 years ago

aLMonotonous commented 4 years ago

Thanks for the brilliant work and for sharing the code.

here I have got confused about the data set size. I followed the guide of IMP but got a dataset of (~50k,~20k) rather than (7.5k,3.2k) which was used by most papers. I checked some issues in the repo, their logs show the same data size of me. So, I'm wondering if I have missed anything?

tks a lot

sangminwoo commented 4 years ago

What do you mean by 50k, 20k? Can you elaborate more?

aLMonotonous commented 4 years ago

What do you mean by 50k, 20k? Can you elaborate more? @sangminwoo I mean the size of (train test) data set.

btw I'm reading your work too, you made great progress. any plan for sharing the code?

jwyang commented 4 years ago

Hi, @aLMonotonous , there are indeed multiple version of data splits. Some of them are for visual relationship detection instead of scene graph generation. For comparison and benchmarking, I followed the most canonical and popular setting in papers like IMP, Neural Motif and our own paper. This is mainly for the fair comparison.