Considerable progress has been made in digitizing realistic and articulated human characters. However, efficiently creating visually plausible biped cartoon characters remains demanding and challenging, mainly due to the lack of data. In this work, we propose to fill this gap by introducing 3DBiCar, the first large-scale full-body 3D biped character data. We build 3DBiCar following three rules:
3DBiCar spans a wide range of 3D biped cartoon characters, containing 1,500 high-quality 3D models. We firstly carefully collect images of 2D full-body biped cartoon characters with diverse identities, shape, and textural styles from the Internet, resulting in 15 character species and 4 image styles. Then we recruit six professional artists to create 3D corresponding character models according to the collected reference images. The modeling result is required to be matched with the reference images as much as possible.
The key to building a linear parametric shape model is keeping a unified mesh topology. Traditional human parametric models utilize a template mesh to register different human body scans with 3D landmarks to keep topologically uniform. Inspired by this, we first create a template mesh with several 3D colored landmarks. All six artists are required to craft 3D models by deforming the above-predefined template under the constraints of these obvious landmarks. We set up a review committee of 10 to check these models based on the landmarks, ensuring the consistency of mesh topology. The landmarks could also be used to compute the position of models' joints for body posing or character animation. The topological consistency of 3DBiCar paves the way to learn a skinned parametric model (e.g., RaBit).
We provide complete and various forms of data for each character. There are not only the 3D shape meshes and UV-space textures carefully crafted by artists but also collected reference images. For each character, artists are asked first to create a T-pose mesh and then deform it to match the reference pose. Furthermore, all the models are rigged and skinned by the predefined skeleton and skinning weight matrix, which supports further animation production of characters. Note that eyeball meshes are extra modeled to support the facial expression in the future better. Each body mesh comprises 38,726 vertices and 77,448 faces, while each eyeball contains 1,025 vertices and 2,046 faces.
Our dataset is available at OneDrive.