Humans can perceive 3D world from images that aren't 3D consistent, but why can't machines? This project enables 3D reconstruction of non-geometrically consistent scenes, such as cartoons. _We take a few images from a cartoon and reconstruct it in 3D. Our project page is at https://toon3d.studio/.

Image 1	Image 2	Image 3

Create a conda environment and install Toon3D.

pip install -e .

Download the Toon3D Dataset from here and place the dataset folders in data/processed.

This section walks you through starting from a small collection of images, annotating them, and running our custom SfM to create the initial dataset. This is how we created the pre-processed datasets, available above.

You can download one of our image collections from this Google Drive link or simply use your own. Place the images in a folder within the parent directory such as data/images/[dataset], where [dataset] is the dataset name. To get started, download the bobs-burgers-dining images and follow the instructions below.

This step runs depth estimation and Segment Anything (SAM).

Install with

pip install segment_anything @ git+https://github.com/facebookresearch/segment-anything.git

Download SAM weights with tnd-download-data sam.

Now, you can run

tnd-process-data initialize --dataset [dataset] --input_path [input_path]

Where the input_path is the source folder of your images while dataset is the name of the dataset that will be output to data/processed/[dataset]

For example,

tnd-process-data initialize --dataset bobs-burgers-dining --input_path data/images/bobs-burgers-dining

If you want to add more images to your dataset after initializing it, run this

tnd-process-data add --dataset bobs-burgers-dining --input_path data/images/more-bobs-burgers-dining-photos

You can use labeler.toon3d.studio.

You have two ways of getting your data in the viewer.

(1) Uploading ZIP folders

You can either upload a zipped process-data folder or simply a selection of images. If just images, you'll have more limited functionality (no depth images or SAM masks).

(2) No-upload version for faster development

Expose your processed data to a public URL. Here is an example with our script with CORS allowed for any origin. You can change the port or the relative directory from which to host the files from.

tnd-server --path data/processed --port 8000

Now, open your processed dataset and annotate.

Navigate to https://labeler.toon3d.studio/?path=http://localhost:8000/[dataset] in this case.

For example: https://labeler.toon3d.studio/?path=http://localhost:8000/bobs-burgers-dining

Now we can run our method to get a dense 3D reconstruction for novel-view synthesis.

For example,

tnd-run --dataset bobs-burgers-dining

For example,

ns-train toon3d --data data/nerfstudio/bobs-burgers-dining

Render a camera path that you created

tnd-render camera-path --load-config [load-config] --camera-path-filename [camera-path-filename] --output-path [output-path].mp4

Render videos between training views

tnd-render interpolate --load-config [load-config] --output-path [output-path]

The outputs folder is organized according to the types of experiments conducted.

outputs/[dataset]/run/[timestamp]       # For SfM experiments
outputs/[dataset]/toon3d/[timestamp]    # For MVS experiments

If you find this code or data useful for your research, please consider citing the following paper:

@inproceedings{weber2023toon3d,
title = {Toon3D: Seeing Cartoons from a New Perspective},
author = {Ethan Weber* and Riley Peterlinz* and Rohan Mathur and
    Frederik Warburg and Alexei A. Efros and Angjoo Kanazawa},
booktitle = {arXiv},
year = {2024},
}