Model checkpoints trained on DocLayNet

[ninikolov]

Hello,

Could you please make some/all of the models you trained for the DocLayNet paper available for download? That would make it really easy to quickly test the utility of DocLayNet, and will make any future comparisons easier.

Thanks!

[cau-git]

Hello @ninikolov, we are glad to hear you would like to check the utility of DocLayNet.

We are not planning to provide the trained model checkpoints as a maintained asset at this time, but it should be very easy to reproduce any object-detection model with this dataset on the latest deep-learning frameworks.

If you want to assess the layout segmentation performance we achieve with DocLayNet, please be invited to check out our public document conversion service in DS4SD here. You can use it from the Web or through our python SDK. We also provide example codes to visualize the output.

[opyate]

but it should be very easy to reproduce any object-detection model with this dataset on the latest deep-learning frameworks.

Hello, I'd be using detectron2, but which architecture do you recommend? E.g. faster_rcnn_R_50_FPN_3x, faster_rcnn_X-101-64x4d-FPN_1x, etc.