Question about Region-wise Quality

[sxj1215]

Thanks for your great work!. I tested the results on LayoutSAM-Eval using the following script:
python train.py --cfg project/plangen/cfg/uni/h_text_ump+oimsam.py --opt test=True resume=./checkpoints/plan/checkpoint-200000 test_data.data_name='creati' test_data.task_type='uni'
While the results on Global-wise Quality match the reported numbers in your paper, the Region-wise Quality appears significantly lower. Specifically, I obtained:

Total number of bbox: 19133
Total score of spatial: 11990.0; Average score of spatial: 0.6267
Total score of color: 10256.0; Average score of color: 0.536
Total score of texture: 10662.0; Average score of texture: 0.5573
Total score of shape: 10407.0; Average score of shape: 0.5439

I wonder if there are any implementation details I may have missed that would affect this evaluation. Also, do you have any plans to release the evaluation code? Many thanks!

[hrz2000]

Thank you for your interest in this work! Have you observed whether the saved image visualization is correct? It's like this:

In addition, are the parameters loaded correctly before inference? Under normal circumstances, only gen_vision_model below are missing.

[sxj1215]

Many thanks for your reply and clarification! BTW, could it be possible to share the grit_json (grit-20m-512-obj3-demo.json), which is used for training PlanGen? Thanks in advance!

[hrz2000]

PlanGen doesn't do much more with the dataset, you can refer to https://github.com/360CVGroup/HiCo_T2I.git to configure the dataset with grit set up~