---
language:
  - en
license: mit
license_link: LICENSE
pretty_name: omnicvproject
tags:
  - 3d
  - timeseries
task_categories:
  - robotics
  - depth-estimation
dataset_info:
  features:
  - name: metadata
    dtype: string
  - name: left
    dtype: image
  - name: right
    dtype: image
  - name: depth
    dtype: image
  splits:
  - name: train
    num_bytes: 901589471
    num_examples: 100
  download_size: 900986407
  dataset_size: 901589471
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Omni Instrument CV Project Dataset

The **Omni Instrument CV Project** is a compact robotics dataset combining:

- [x] Raw ROS 2 MCAP bag (stereo, depth, TF, odom)  
- [x] Ground truth mesh (STL)  
- [x] Distilled synchronized dataset (100 samples)  

---

# Contents

## 1. **Raw Data**

Contains a stripped ROS 2 MCAP recording:

- `/zed/zedxm/left/color/rect/image`
- `/zed/zedxm/right/color/rect/image`
- `/zed/zedxm/depth/depth_registered`
- `/zed/zedxm/odom`
- `/tf`, `/tf_static`
- Associated `CameraInfo`

<p align="center">
  <img src="assets/neural_depth_demo.gif" width="1000" style="object-fit:fill;">
</p>

## 2. **Meshes**

Includes:

- [omni_mesh.stl](meshes/omni_mesh.stl) — ground-truth geometry for TSDF + reconstruction metrics.

<p align="center">
  <img src="assets/ground_truth_mesh.png" width="1000" style="object-fit:fill;">
</p>

## 3. **HuggingFace Sample**

Load a sample and decode its metadata:

```python
from datasets import load_dataset
import json

ds = load_dataset("OmniInstrument/CV_project")

sample = ds[0]

meta = json.loads(sample["metadata"])

timestamp = meta["timestamp"]
position  = meta["position"]        # [x, y, z]
orientation = meta["orientation"]   # [qx, qy, qz, qw]

print("Timestamp:", timestamp)
print("Position:", position)
print("Orientation:", orientation)
```

## License
This software and dataset are released under the [MIT License](LICENSE).