Omnidata (Steerable Datasets)

A Scalable Pipeline for Making Multi-Task Mid-Level Vision Datasets from 3D Scans (ICCV 2021)

Project Website · Paper · >> [Github] << · Data · Pretrained Weights · Annotator ·

Pretrained models

We provide huggingface demos for monocular surface normal estimation and depth estimation. You can load/run the models

import torch
# you may need to install timm for the DPT (we use 0.4.12)

# Surface normal estimation model
model_normal = torch.hub.load('alexsax/omnidata_models', 'surface_normal_dpt_hybrid_384')

# Depth estimation model
model_depth = torch.hub.load('alexsax/omnidata_models', 'depth_dpt_hybrid_384')

# Without pre-trained weights
model_custom = torch.hub.load('alexsax/omnidata_models', 'dpt_hybrid_384', pretrained=False, task='normal')

Previously, installing + using the models was more difficult. Using torch.hub.load is now the recommended way to use the models locally.

---

Table of Contents

• Pretrained models
  • online demo [Updated Transormer models - Mar 2022]
  • weights and pytorch demo [Updated Transormer models - Mar 2022]
  • training
• Data
  • dataset info
  • download standardized portions of the dataset [Added Hypersim fix - Aug 2022]
  • configurable pytorch dataloaders [Added multiview, mesh, camera parameters - Aug 2022]
• The Pipeline
  • generating more multi-view data
  • online steerable dataset demo
• Source for all above
• Paper code (#MiDaS loss)
• Citing

---

Demo code, training losses, etc are available here: weights and code:

python demo.py --task depth --img_path $PATH_TO_IMAGE_OR_FOLDER --output_path $PATH_TO_SAVE_OUTPUT    # or TASK=normal

---

Dataset

You can download each component and modality individually or all at once with our download utility. MAIN DATA PAGE

conda install -c conda-forge aria2
pip install 'omnidata-tools'

omnitools.download point_info rgb depth_euclidean mask_valid fragments \
    --components replica taskonomy \
    --subset debug \
    --dest ./omnidata_starter_dataset/ \
    --name YOUR_NAME --email YOUR_EMAIL --agree_all

We ran our annotation pipeline on several collections of 3D meshes. The result is a 24-million-viewpoint multiview dataset comprising over 2000 scenes with the following labels for each image:

Per-Image Information

RGB	Cam. Intrinsics	Cam. Pose	Correspondences (Flow)	Segm. (Instances)	Segm. (Semantic)	Segm. (2D Graphcut)	Segm. (2.5D Graphcut)
Distance (Euclidean)	Depth (Z-Buffer)	Surface Normals	Curvature	Edges (Texture)	Shading (reshading)	Keypoints (2D, SIFT)	Keypoints (3D, NARF)
Masks (valid pixels)	Shading

--components: Taskonomy, Hypersim, Replica, Google Scanned Objects in Replica, Habitat-Matterport3D, BlendedMVS, CLEVR

More about the data: Standardized data subsets and download tool

---

Annotate a new 3D mesh

git clone https://github.com/Ainaz99/omnidata-annotator # Generation scripts
docker pull ainaz99/omnidata-annotator:latest           # Includes Blender, Meshlab, other libs
docker run -ti --rm \
   -v omnidata-annotator:/annotator \
   -v PATH_TO_3D_MODEL:/model \
   ainaz99/omnidata-annotator:latest
cd /annotator
./run-demo.sh

Documentation and a tutorial here.

Source code

git clone https://github.com/EPFL-VILAB/omnidata
cd omnidata_tools/torch # PyTorch code for configurable Omnidata dataloaders, scripts for training, demo of trained models
cd omnidata_tools       # Code for downloader utility above, what's installed by: `pip install 'omnidata-tools'`
cd omnidata_annotator   # Annotator code. Docker CLI above
cd paper_code           # Reference

Citing

@inproceedings{eftekhar2021omnidata,
  title={Omnidata: A Scalable Pipeline for Making Multi-Task Mid-Level Vision Datasets From 3D Scans},
  author={Eftekhar, Ainaz and Sax, Alexander and Malik, Jitendra and Zamir, Amir},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={10786--10796},
  year={2021}
}

In case you use our latest pretrained models please also cite the following paper for 3D data augmentations:

@inproceedings{kar20223d,
  title={3D Common Corruptions and Data Augmentation},
  author={Kar, O{\u{g}}uzhan Fatih and Yeo, Teresa and Atanov, Andrei and Zamir, Amir},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={18963--18974},
  year={2022}
}

...were you looking for the research paper or project website?