TEDs

[cqray1990]

i use text dect model as you supply,get PSENet detection results, TableMASTER structure recognition results and MASTER textline recognition results, and use match.py to get TEDS score. PSENet + MASTER + TableMASTER_maxlength_500: In the validation set, the accuracy is 0.56. also is not 0.9658

i do valiation step by step as readme refer to, can't not recurrent the results: here is my cfg in table_inference.py:

cfg = {
    'pse_config':'/configs/textdet/psenet/psenet_r50_fpnf_600e_pubtabnet.py',
    'master_config':'/configs/textrecog/master/master_lmdb_ResnetExtra_tableRec_dataset_dynamic_mmfp16.py',
    'structure_master_config':/configs/textrecog/master/table_master_lmdb_ResnetExtract_Ranger_0930.py',
    'pse_ckpt':'./pse_epoch_600.pth',
    'master_ckpt':'./master_epoch_6.pth',
    'structure_master_ckpt':'./epoch_16_0.7767.pth',
    'end2end_result_folder':'/mmocr_pubtabnet_recognition_0917val/work_dirs/end2end_val_result',
    'structure_master_result_folder':'/mmocr_pubtabnet_recognition_0917val/work_dirs/structure_val_result',

    'test_folder':'/pubtabnet/pubtabnet/val',
    # 'test_folder':'./smallVal10'
    'chunks_nums':chunk_nums
}

[delveintodetail]

Visualize result of each step. Check if the text line detection model, the accuracy of the table structure recognition, the text line detection are all right.

[JiaquanYe]

i use text dect model as you supply,get PSENet detection results, TableMASTER structure recognition results and MASTER textline recognition results, and use match.py to get TEDS score. PSENet + MASTER + TableMASTER_maxlength_500: In the validation set, the accuracy is 0.56. also is not 0.9658

i do valiation step by step as readme refer to, can't not recurrent the results: here is my cfg in table_inference.py:

cfg = {
    'pse_config':'/configs/textdet/psenet/psenet_r50_fpnf_600e_pubtabnet.py',
    'master_config':'/configs/textrecog/master/master_lmdb_ResnetExtra_tableRec_dataset_dynamic_mmfp16.py',
    'structure_master_config':/configs/textrecog/master/table_master_lmdb_ResnetExtract_Ranger_0930.py',
    'pse_ckpt':'./pse_epoch_600.pth',
    'master_ckpt':'./master_epoch_6.pth',
    'structure_master_ckpt':'./epoch_16_0.7767.pth',
    'end2end_result_folder':'/mmocr_pubtabnet_recognition_0917val/work_dirs/end2end_val_result',
    'structure_master_result_folder':'/mmocr_pubtabnet_recognition_0917val/work_dirs/structure_val_result',

    'test_folder':'/pubtabnet/pubtabnet/val',
    # 'test_folder':'./smallVal10'
    'chunks_nums':chunk_nums
}

We have released the TEDS evaluation log, you can download the log file by this link, and check each val image's TEDS score.

[cqray1990]

@JiaquanYe i find detect is right ,but text recon is wired

[JiaquanYe]

@JiaquanYe i find detect is right ,but text recon is wired

Maybe your alphabet is different with the text-line pretrained model. We have released the alphabet in “table_recognition/demo/alphabet”