Awesome
PubLayNet
PubLayNet is a large dataset of document images, of which the layout is annotated with both bounding boxes and polygonal segmentations. For more information, see PubLayNet original
| <img src="./example_images/PMC4334925_00006.jpg" width=400> | <img src="./example_images/PMC538274_00004.jpg" width=400> |
|---|---|
| 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)
<img src="https://user-images.githubusercontent.com/24642166/75600546-066b6900-5ae3-11ea-9774-a0a0396e6fb1.png" width=1000>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