thomas0809 / MolScribe

Robust Molecular Structure Recognition with Image-to-Graph Generation
MIT License
153 stars 31 forks source link

感谢你开源的MolScribe模型,已经试用了,准确性还不错。想请教一下,关于分子位置检测有推荐的开源模型吗? #9

Closed SnailWhb closed 8 months ago

LingjieBao1998 commented 1 year ago

刚好最近也在follow这些工作,这篇工作不错(https://github.com/Kohulan/DECIMER-Image-Segmentation),不过召回率的话不是100%。当然还是很感谢MolScribe模型的开源工作。

SnailWhb commented 1 year ago

刚好最近也在follow这些工作,这篇工作不错(https://github.com/Kohulan/DECIMER-Image-Segmentation),不过召回率的话不是100%。当然还是很感谢MolScribe模型的开源工作。

谢谢回复,DECIMER-IMAGE-Segmentation召回率有待提高,它是基于Mask-RCNN训练的,想要基于Transformer框架重新训练一个分子检测模型,但无奈没有找到相关的公开数据集.

LingjieBao1998 commented 1 year ago

刚好最近也在follow这些工作,这篇工作不错(https://github.com/Kohulan/DECIMER-Image-Segmentation),不过召回率的话不是100%。当然还是很感谢MolScribe模型的开源工作。

谢谢回复,DECIMER-IMAGE-Segmentation召回率有待提高,它是基于Mask-RCNN训练的,想要基于Transformer框架重新训练一个分子检测模型,但无奈没有找到相关的公开数据集.

是的,他们没有提供数据集,但是原始文献中文章中提供了数据标注的方式。 如果你只是拿他们的模型来用的话,简单的方式可以通过稍微旋转图片来提高召回率(大致的召回率能够达到80%以上)。

thomas0809 commented 1 year ago

谢谢你对我们工作的兴趣!

我们有一个Transformer-based molecule detection model,最近(可能一两周)就会公开,我过几天在这里更新。

thomas0809 commented 1 year ago

Hi,

I would like to introduce you to a new tool we have developed: https://github.com/Ozymandias314/MolDetect. It is a Transformer-based molecule detection model. We will be releasing our manuscript soon. Please stay tuned.

LingjieBao1998 commented 1 year ago

Hi,

I would like to introduce you to a new tool we have developed: https://github.com/Ozymandias314/MolDetect. It is a Transformer-based molecule detection model. We will be releasing our manuscript soon. Please stay tuned.↳

Thanks for your hard work. Could you release the link or DOI of paper.

SnailWhb commented 1 year ago

Hi,

I would like to introduce you to a new tool we have developed: https://github.com/Ozymandias314/MolDetect. It is a Transformer-based molecule detection model. We will be releasing our manuscript soon. Please stay tuned.

非常感谢,已测试。整体效果相对DECIMER-Image-Segmentation要好一些,但对于分子之间距离很小时,出错的概率较大。 image

image