PubLayNet

PubLayNet is a large dataset of document images, of which the layout is annotated with both bounding boxes and polygonal segmentations.

PMC4334925_00006 PMC538274_00004
PMC4334925_00006.jpg PMC538274_00004.jpg

Recent updates

15/Sept/2020 - Add training code.

29/Feb/2020 - Add benchmarking for maskrcnn_resnet50_fpn.

22/Feb/2020 - Pre-trained Mask-RCNN model in (Pytorch) are released .

Benchmarking

Architecture Iter num (x16) AP AP50 AP75 AP Small AP Medium AP Large MD5SUM
MaskRCNN-Resnet50-FPN 196k 0.91 0.98 0.96 0.41 0.76 0.95 393e6700095a673065fcecf5e8f264f7

Demo

Download trained weights in Benchmarking section above, locate it in maskrcnn directory

Run

cd maskrcnn
python infer.py --image_path = "document_image_dir/image.jpg" --model_path = "mrcnn_model_dir/model.pth" --output_path="model_segmentation_output_dir/"

Avarage Precision in validation stages (via Tensorboard)

75600546-066b6900-5ae3-11ea-9774-a0a0396e6fb1

Training

Please take a look at training_code dir. Sorry for the dirty code but I really don't have time to refactor it :D

GitHub