search

mxin262 / ESTextSpotter

(ICCV 2023) ESTextSpotter: Towards Better Scene Text Spotting with Explicit Synergy in Transformer
72 stars 7 forks source link

Evaluation Scores for CTW Recognition ‘Full’ #11

Open ZhengyaoFang opened 10 months ago

ZhengyaoFang commented 10 months ago

Hello,

I've been running the provided code to conduct recognition evaluation for CTW. I noticed that the score for 'None' matches the one listed on your GitHub (64.9999), which is great. However, when I test for 'Full', the score I obtained is only 77.33, which seems to be lower than expected.

I would like to confirm if the only changes needed to test 'Full' are setting use_lexicon=True and lexicon_type=1 in coco_transformer.py. Could there be any other settings or steps that I might be missing to achieve the expected performance?

Thank you for your assistance and for sharing your work.

ZhengyaoFang commented 10 months ago

bash scripts/test.sh config/ESTS/ESTS_5scale_ctw1500_finetune.py datasets 1 ctw1500_checkpoint.pth ctw1500_val

mxin262 commented 10 months ago

Hi, you can try to adjust the score to evaluate on the 'Full' lexicon.https://github.com/mxin262/ESTextSpotter/blob/21717310bfb9b32108e5124206dcaf66d90a5e34/evaluation/text_evaluation_all.py#L71