|
3D Vision and Embodied AI |
|
| THuman-Sit and THuman-Grasp dataset |
| THuman-Sit and THuman-Grasp datasets contains 4,700 and 500 high-quality scans, respectively, captured by a dense DSLR rig. These datasets capture a wide range of human-chair and hand-object interactions, offering a diverse collection of sitting and grasping poses. For each scan, we provide the 3D model (****.obj) along with the corresponding texture map. We also provide the fitting results of parametric model, like SMPL, MANO and NIMBLE. |
|
|
| THuman3.0 Dataset |
| THuman3.0 Dataset contains 20 combinations of human-garment, each contains 15 to 35 high-quality human scans captured by a dense DSLR rig(128 cameras). For each scan, we provide the 3D model (****.obj) along with the corresponding texture map. |
|
|
| THuman2.0 Dataset |
| THuman2.0 Dataset contains 500 high-quality human scans captured by a dense DSLR rig (128 cameras). For each scan, we provide the 3D model (****.obj) along with the corresponding texture map. |
|
|
| MultiHuman Dataset |
| MultiHuman dataset consists of 150 human scans captured by a dense DSLR rig, which is used to evaluate multi-person reconstruction algorithms. |
| For each scene, we provide the normalized 3D model (obj_all/***), the cropped single person 3D models (obj/***), and the estimated SMPLX models (smplx/***). The SMPLX models are fitted from 6 views, which can be used as baselines or ground truth. |
|
|
| THUman Dataset |
| We provide the SMPL mesh models (smpl.obj) as well as the the corresponding shape&pose; parameters (smpl_params.txt) in our dataset. |
|
|
| Thuman5.0 Dataset |
| THuman5.0 Dataset contains 10 multi-view video sequences used in our paper "DiffuStereo: High Quality Human Reconstruction via Diffusion-based Stereo Using Sparse Cameras". They are captured with 32 well-calibrated RGB cameras in 30 fps, with a resolution of 4096x3000 and lengths ranging from 1000 to 5000 frames. |
|
|
| THuman4.0 Dataset |
| THuman4.0 Dataset contains three multi-view image sequences used in our paper "Structured Local Radiance Fields for Human Avatar Modeling". They are captured with 24 well-calibrated RGB cameras in 30 fps, with a resolution of 1330x1150 and lengths ranging from 2500 to 5000 frames. We also provide the SMPL fitting in the dataset. |
|
|
| FaceVerse Dataset |
| FaceVerse-High Quality 3D Face Dataset contains 2,688 high-quality head scans (21 expressions from 128 identities) captured by a dense DSLR rig. For each scan, we provide the 3D model (.obj), the corresponding texture map (.jpeg) and the FaceVerse fitted model (.ply) with the same topology. |
|
|
