README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - animal
size_categories:
  - n<1K

External Data for 4D-Animal

External data from the paper: 4D-Animal: Freely Reconstructing Animatable 3D Animals from Videos.

| ArXiv | Code |

Data Tree

.
├── cse                  # DensePose model weights and configuration files
│   ├── Base-DensePose-RCNN-FPN.yaml
│   ├── densepose_rcnn_R_50_FPN_soft_animals_I0_finetune_i2m_16k.yaml
│   └── model_final_8c9d99.pkl
├── lbos                 # Laplace-Beltrami Operator results
│   ├── lbo_cse.pk
│   ├── lbo_cse_to_smal.pk
│   └── lbo_smal.pk
├── refined_masks        # Refined segmentation masks
├── preprocessing
│   ├── cse_embeddings   # Continuous Surface Embedding (CSE) data
│   └── pnp_ransac       # Refined camera parameters via PnP + RANSAC
├── smal                 # SMAL model data
│   ├── my_smpl_39dogsnorm_newv3_dog.pkl
│   └── symmetry_inds.json
├── sparse_keypoints     # Sparse keypoints predicted using BITE
├── textures             # RGB texture for CSE and SMAL
│   ├── texture_cse.pk
│   └── texture_smal.pk
└── cop3d_data           # Subset of COP3D dataset

DensePose Model Weights and Configuration Files

Pretrained DensePose models and configuration files from Detectron2:

• densepose_rcnn_R_50_FPN_soft_animals_I0_finetune_i2m_16k.yaml
• Base-DensePose-RCNN-FPN.yaml
• model_final_8c9d99.pkl

Laplace-Beltrami Operators

For details on Laplace-Beltrami operators used in this project, refer to Section 3.1 of Continuous Surface Embeddings (CSE) paper.

SMAL Model

The SMAL model files are taken from the project BITE: Beyond Priors for Improved 3D Dog Pose Estimation:

• my_smpl_39dogsnorm_newv3_dog.pkl
• symmetry_inds.json

Sparse Keypoints

Sparse keypoints are generated using the keypoint prediction module provided by BITE.

COP3D Dataset

This project is built using the COP3D dataset. We provide a subset of 50 videos used in our experiments. To access the full dataset, follow the official download instructions.

Each video directory contains:

• images: Original video frames from COP3D
• masks: Corresponding foreground masks
• part_masks: Part segmentation masks (head, tail, body, legs) generated using PartGLEE
• tracking_points.pt: Tracking data generated using BootsTAP

License

This is not a new dataset. All data used in this project are derived from existing public sources, such as COP3D, BITE, and others.

We fully adhere to the terms of use and licenses of all referenced datasets. If you believe that any part of the data presented here violates applicable rights or policies, please contact us. We are committed to addressing any such concerns promptly and appropriately.

Citation

If you find our work useful, please consider citing:

@article{zhong20254danimal,
  title={4D-Animal: Freely Reconstructing Animatable 3D Animals from Videos},
  author={Zhong, Shanshan and Peng, Jiawei and Zheng, Zehan and Huang, Zhongzhan and Ma, Wufei and Zhang, Guofeng and Liu, Qihao and Yuille, Alan and Chen, Jieneng},
  journal={ArXiv},
  year={2025},
  volume={abs/2507.10437},
}