Awesome

DocIIW

Repository for the paper "Intrinsic Decomposition of Document Images In-the-Wild" (BMVC '20)

Quick Links: PDF | arXiv | Talk | Supplementary

Updates

• Sep 5th, 2020: Initial data is released (90K images).
• Mar 20th, 2021: Evaluation images are released.
• Nov 8th, 2022: Training Code and models.
• Coming Soon: Training details.

Doc3DShade

Doc3DShade extends Doc3D with realistic lighting and shading. Follows a similar synthetic rendering procedure using captured document 3D shapes but final image generation step combines real shading of different types of paper materials under numerous illumination conditions. <br> Following figure illustrates the image generation pipeline:

Following figure shows a side-by-side comparison of images in Doc3DShade and Doc3D:

Data Download Instructions

Doc3Dshade contains 90K images, 80K used for training and 10K for validation. Split used in the paper: train, val

• Download the input images from img.zip .
• Download the white-balanced images from wbl.zip .
• Download synthetic textures from alb.zip .

Training Instructions

• Upcoming

Pre-trained Models

• All models: GDrive Link
• WBNet: GDrive Link
• SMTNet: GDrive Link
• SMTNet(w/ adversarial loss): GDrive Link

Evaluation Images and Results

• Real test images are given in: /testimgs/real
• Shading removed real test images:
  • Basic: GDrive Link
• Shading removed DocUNet [1] images are available at:
  • Basic: GDrive Link
  • With adversarial loss: GDrive Link
• Shading removed and unwarped [2] DocUNet [1] images are available at:
  • Basic: GDrive Link
  • With adversarial loss: GDrive Link

Citation:

If you use the dataset, please consider citing our work-

@inproceedings{DasDocIIW20,
  author    = {Sagnik Das, Hassan Ahmed Sial, Ke Ma, Ramon Baldrich, Maria Vanrell and Dimitris Samaras},
  title     = {Intrinsic Decomposition of Document Images In-the-Wild},
  booktitle = {31st British Machine Vision Conference 2020, {BMVC} 2020, Manchester, UK, September 7-10, 2020},
  publisher = {{BMVA} Press},
  year      = {2020},
}

References:

[1] DocUNet: https://www3.cs.stonybrook.edu/~cvl/docunet.html

[2] DewarpNet: https://sagniklp.github.io/dewarpnet-webpage/