the terrible recognize result with fine-tunning abcnet v1 using TotalText datasets

[dreamlychina]

thanks for sharing this amazing work. I did not change any code and used the pretrained checkpoint pretrain_attn_R_50.pth to fine-tune the abcnetv1 using totaltext datasets, the detection results seems normal, but the recognize result is terrible. [11/18 17:08:14 d2.engine.defaults]: Evaluation results for totaltext_val in csv format: [11/18 17:08:14 d2.evaluation.testing]: copypaste: Task: DETECTION_ONLY_RESULTS [11/18 17:08:14 d2.evaluation.testing]: copypaste: precision,recall,hmean [11/18 17:08:14 d2.evaluation.testing]: copypaste: 0.8692,0.8342,0.8513 [11/18 17:08:14 d2.evaluation.testing]: copypaste: Task: None-E2E_RESULTS [11/18 17:08:14 d2.evaluation.testing]: copypaste: precision,recall,hmean [11/18 17:08:14 d2.evaluation.testing]: copypaste: 0.0000,0.0000,0.0000 [11/18 17:08:14 d2.evaluation.testing]: copypaste: Task: Strong-E2E_RESULTS [11/18 17:08:14 d2.evaluation.testing]: copypaste: precision,recall,hmean [11/18 17:08:14 d2.evaluation.testing]: copypaste: 0.0022,0.0022,0.0022

[EternalSonata]

I also encountered the same problem, have you solved the problem?

[dreamlychina]

I also encountered the same problem, have you solved the problem?

I have not solved the problem，if you do， please tell me.

[EternalSonata]

I also encountered the same problem, have you solved the problem?

I have not solved the problem，if you do， please tell me.

Sorry, I haven't figured it out yet, but I have a feeling the problem is related to recognition branches... As I applied the final model to the dataset and printed the images, I found that only the text areas were detected on the image, but the text was not recognized.