ying09 / TextFuseNet

A PyTorch implementation of "TextFuseNet: Scene Text Detection with Richer Fused Features".
MIT License
476 stars 123 forks source link

fail to reproduce ICDAR2015 result #111

Open D641593 opened 1 year ago

D641593 commented 1 year ago

Hi, thank you for a great work.

I'm having some trouble reproducing the ICDAR2015 result.

First, I use the synth-text pretrained weight to generate character labels.

According to #99 , I fill the boxes with confidence scores > 0.8 to generate masks.

Then, I start training using synth-text pretrained weights and the generated COCO format label.

The result of the model is {"precision": 0.7594320486815416, "recall": 0.9012999518536351, "hmean": 0.8243064729194188}.

Although Recall is very close to the original data on paper, Precision drops a lot.

Could anyone give me some help?