jz462 / Large-Scale-VRD.pytorch

Implementation for the AAAI2019 paper "Large-scale Visual Relationship Understanding"
https://arxiv.org/abs/1804.10660
MIT License
144 stars 25 forks source link

strange BBOX format in relationship annotation #13

Closed czqInNanjing closed 4 years ago

czqInNanjing commented 5 years ago

Hello, thank you for your great job! It is very helpful.

I just wonder why the bbox format of the subject and object in relationship annotation is y1y2x1x2 rather than xywh or x1y1x2y2, which makes me very confuse.

I spend sometime fixing this bug since it is sometimes very difficult to notice small errors that seems to be right.

I think it would be better to highlight this special thing in the readme.

jz462 commented 5 years ago

Hi @czqInNanjing,

The y1y2x1x2 format is kept merely for the reason that when I was doing the project I needed to take care of OpenImages, Visual Genome and VRD datasets at the same time, and I chose this one to be consistent across the three datasets. I strongly suggest not to change the default format in the code but instead modify your dataset to fit in this format, which is definitely less risky.

Ji