doc3D

December 18, 2025 ยท View on GitHub

Doc3D is the first 3D dataset focused on document unwarping with realistic paper warping and renderings.

It contains 100k images with the following ground-truths:

  • 3D Coordinates
  • Depth
  • UV
  • Backward Mapping
  • Albedo
  • Normals
  • Checkerboard

Doc3D dataset is now available through HuggingFace

Download instructions (Will be deprecated by 31st December, 2025)

This repository contains all the necessary bash scripts to download the dataset-

  • To download the dataset you need to obtain a username and password. Please fill out the Google Form to request one.
  • Update the assigned username password in the download scripts at lines: local uname=**** local pass=****
  • To download the entire dataset at once (in the default directory $HOME/Downloads/doc3d), use the following command:
    • bash download_doc3d.sh
  • To download in a specific directory-
    • bash download_doc3d.sh <out_dir>
  • Individual bash scripts are provided to download a specific part of the data. Following will download all the image files in <out_dir>/doc3d/img/-
    • bash download_img.sh <out_dir>

Some Notes:

  • A download can be interrupted and resumed later, wget keeps track of it.
  • Already downloaded files will be skipped and partially downloaded files will be resumed.
  • The scripts are tested on Linux and Mac. For windows, a bash shell [probably-useful] should work.

Train and test split: should be done based on the mesh IDs, follow here: https://github.com/cvlab-stonybrook/doc3D-dataset/issues/8#issuecomment-546566941

Rendering codes are available!!: You can use the scripts here to render your own version of doc3D.

Visualize Data:

Run the demo.py file to display a random image and corresponding ground-truths. demo.py takes the following flags-

  • --data_root : Path to the doc3d dataset.
  • --folder : Specific folder to load image from.
  • --download_sample : If you want to download some samples and run demo.py on it. useful if you want to visualize it before downloading the entire data.
  • --unwarp : Unwarp input image using the ground-truth backward mapping.

Release Updates:

  • Sep 16, 2019: v0.5 (36K images, no depth map)
  • Sep 17, 2019: v0.5.1 (Depth maps for v0.5 images)
  • Sep 21, 2019: Rendering code is now available!
  • Sep 22, 2019: v0.9 (65K images, no albedos)
  • Mar 11, 2020: Please send an email to request the meshes (.obj)
  • Dec 18, 2025: Dataset is available in HuggingFace

Citation:

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

@inproceedings{SagnikKeICCV2019, 
Author = {Sagnik Das*, Ke Ma*, Zhixin Shu, Dimitris Samaras, Roy Shilkrot}, 
Booktitle = {Proceedings of International Conference on Computer Vision}, 
Title = {DewarpNet: Single-Image Document Unwarping With Stacked 3D and 2D Regression Networks}, 
Year = {2019}}

Acknowlegement: