UPDATE

  • 2022.09.16: DART got accepted by NeurIPS 2022 - Datasets and Benchmarks Track!
  • 2022.09.29: DART's GUI source code publicly available at Unity GUI source code!

Abstract

Hand, the bearer of human productivity and intelligence, is receiving much attention due to the recent fever of 3D digital avatars. Among different hand morphable models, MANO has been widely used in various vision & graphics tasks. However, MANO disregards textures and accessories, which largely limits its power to synthesize photorealistic & lifestyle hand data. In this paper, we extend MANO with more Diverse Accessories and Rich Textures, namely DART. DART is comprised of 325 exquisite hand-crafted texture maps which varies in appearance, and covers different kinds of blemishes, make-ups and accessories. We also provide the Unity GUI which allows people to render hands with user-specific settings, e.g pose, camera, background, lighting, and DART's textures. In this way, we generate large-scale (800K), diverse, and high-fidelity hand images, paired with perfect-aligned 3D labels, called DARTset. Experiments demonstrate its superiority in generalization and diversity. As a great complement for existing datasets, DARTset could boost hand pose estimation & surface reconstruction tasks. DART and Unity software is publicly available for research purpose.


overview

Video



Comparison between DART and MANO basic topology.

In DART, we remould the standard template hand mesh of MANO, which has 778 vertices and 1,538 faces, to a wrist-enhanced template mesh of 842 vertices and 1,666 faces. The pose parameter that drive the template hand mesh in MANO can be used as a direct placement for DART without any modifications.

How to use DART tool in generating your own data?

Step1: click DART GUI & Code and download Build_Hand.zip, unzip and execute Hand.exe.

Step2: Pose Editing: allow arbitrarily, illumination, accessory, skin color and other terms.

Step3: Exporting: rendered image with GT(mano pose, 2d/3d joint are put into the output.pkl)

2022.09.22 We put the postprocess folder into link. Please use that code for postprocessing the intermediate output yielded by DART's GUI!

DARTset Datasheet & Explanation

Dataset Structure

DARTset is composed of train and test. The folder of each is described as below.

Train set
* Train: 729,189 single hand frames, 25% of which are wearing accessories.

Test set
* Test (H): 288,77 single hand frames, 25% of which are wearing accessories.

Total set
* DARTset: 758,066 single hand frames. Noteworthy here, we conduct experiments on full 800K DARTset and filter out ~42,000 of images which left wrist unsealed on the final version.
                                

Pose sampling we use spherical linear interpolation (Slerp) in pose and root rotation sampling. Among these hands, ~25% are assigned an accessory. In DARTset, basic UV map (skin tones, scars, moles, tattoos) and accessories are all uniformly sampled, the number of their corresponding renders are roughly equal.

Dataset Creation

Texture map (4096 x 4096) are all created manually by 3D artists. GUI and batch data generator is programmed by DART's authors.

  • (a) Pose Sampling: In DART, we totally sampled 800K pose, 758K of them are valid. For each pose, we adopt synthetic pose θs from A-MANO and randomply picked 2,000 pose from FreiHand θr, through calculate the difference between θs and θr, we select the one θr_max that differs most from θs. Then we interpolate 8 rotations from θr_max to θs by spherical linear interpolation (Slerp). For more detail, please check the paper.
  • (b) Mesh Reconstruction & Rendering: After part (a), we have N x (16,3) MANO poses (N=758K). We input those poses into our revised MANO layer directly. Through which we can easily get DART mesh (MANO mesh plus extra wrist) with 842 vertices and 1,666 faces.
  • (c) Rendering: We adopt Unity and our revised SSS skin shader for rendering. During the rendering pipeline, we randomly select the illumination (intensity, color and position), texture (including accessory). After rendering, we have the final image and 2D reprojected landmarks.
  • (d) Ground Truth Data: At part (a) and (b), we already gather all 3D related ground truth for monocular 3D pose estimation & hand tracking yet leave 2D landmark untouched. After the rendering process (part (c)), we obtain the 2D landmark and inject it to the ground truth data file. Since our method in creating DARTset is build on top of the ground truth annotations, so the ground truth annotations are born with renderings.

Considerations for Using the Data
  • Social Impact: No ethical issues because DART do not involve any human biological information.
  • Gender Biases: Since our basic mesh topology is FIXED (we focus more on pose rather than shape), which means the mesh itself is gender-agnostic. In our data generation pipeline, both texture maps and accessories in DARTset are randomly sampled. Considering this, we do not think it may involve extreme gender bias in DARTset.

Additional Information

Acknowledgements


The website template was borrowed from Michaël Gharbi.