UniGeo: Taming Video Diffusion for Unified Consistent Geometry Estimation

In this repo, we provide a unified framework for geometry estimation and evaluation.

Our framework provides a convenient interface for various dataset and various methods, which supports a fair comparison by aligning the output and evaluation scripts.

Evaluation (in progressing ... )

The dataset, model and evaluation metric configuration can be set in the yaml file in configs. E.g.,

• Dataset Config:

  dataset: "ScannetPPDataset"
  root: /path/to/dataset
  h: 384
  w: 512
  clip_length: 25
  clip_overlap: 5
  split: "test"
  

• Model Config:

  model_name: "DepthCrafter"
  model_params:
    model_dir: /path/to/model
    unet_path: /path/to/DepthCrafter_unet
    pre_train_path: /path/to/stable_video_diffusion
  

• Metric Config

  eval_depth:
    metric_names: 
      - 'Abs Rel'
      - 'delta < 1.25'
      - 'delta < 1.$25^{2}$'
      - 'delta < 1.$25^{3}$'
    depth_alignment: "lstsq"
  
  eval_normal:
    metric_names: 
      - 'normal mean'
      - 'normal median'
      - 'angle < 7.5'
      - 'angle < 11.25'
  

Finally, the evaluation process can be performed by

  python eval.py

Supported Datasets

Please refer to dataset for more details.

Supported Methods

Please refer to model for more details.

Acknowledgements

This code borrows heavily from Spann3R, Monst3R, CUT3R and robustmvd. Thanks for these awesome works.

---

If you find this work helpful, please consider citing

@article{sun2025unigeo,
  title={UniGeo: Taming Video Diffusion for Unified Consistent Geometry Estimation},
  author={Sun, Yang-Tian and Yu, Xin and Huang, Zehuan and Huang, Yi-Hua and Guo, Yuan-Chen and Yang, Ziyi and Cao, Yan-Pei and Qi, Xiaojuan},
  journal={arXiv preprint arXiv:2505.24521},
  year={2025}
}