Jingkang50 / OpenPSG

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

The Code with mmcv is hard to understand. #74

Closed BJHYZJ closed 1 year ago

BJHYZJ commented 1 year ago

Hello, dear authors. Thank you for your contribution to the field of scene graph generation, I find your work very interesting, but when I was about to reproduce your article or wish to follow your work, I was very uncomfortable with the opaqueness of the mmcv framework and could not modify the code properly, so I would like to ask if you could provide a pure pytorch version of the code or provide us with some reference materials for learning mmcv, I would appreciate it.

Jingkang50 commented 1 year ago

Thanks for the question. The PSG task contains important components from detection and segmentation, which is not easy to be implemented by pure pytorch. MMdet is one of the best detection framework, which would make the code easier. If you want to dive deep in PSG and dense prediction tasks, I believe you should be familiar with the framework. You can use "debug" mode in vscode to explore the mmdet and mmcv frameworks. Good luck!

BJHYZJ commented 1 year ago

Ok, thank you for your reply.