|
|
--- |
|
|
license: apache-2.0 |
|
|
pipeline_tag: depth-estimation |
|
|
tags: |
|
|
- '360' |
|
|
- depth-estimation |
|
|
- computer-vision |
|
|
--- |
|
|
|
|
|
# ORB: Omni-directional Reconstruction Backbone |
|
|
|
|
|
 |
|
|
|
|
|
**ORB** is a 360° panorama depth estimation model that predicts dense distance maps from equirectangular panoramas in a single forward pass. |
|
|
|
|
|
## Model Description |
|
|
|
|
|
This model takes a 360° equirectangular panorama (2:1 aspect ratio) as input and outputs a dense depth/distance map at the same resolution. It's designed for: |
|
|
|
|
|
- **Zero-shot depth estimation** from panoramic images |
|
|
- **Scale-invariant predictions** with geometric fidelity |
|
|
- **End-to-end processing** without post-processing |
|
|
|
|
|
## Quick Start |
|
|
|
|
|
```python |
|
|
from orb import predict_pano_depth |
|
|
|
|
|
# Predict depth from panorama |
|
|
distance = predict_pano_depth('panorama.png') |
|
|
``` |
|
|
|
|
|
## Model Details |
|
|
|
|
|
- **Input**: RGB panorama (equirectangular, width = 2 × height) |
|
|
- **Output**: Dense depth/distance map (same resolution as input) |
|
|
- **Format**: SafeTensors (1.3 GB) |
|
|
- **Precision**: FP32 / FP16 supported |
|
|
- **Base Architecture**: Built upon [DA²: Depth Anything in Any Direction](https://arxiv.org/abs/2509.26618) |
|
|
|
|
|
## 📖 Full Documentation |
|
|
|
|
|
For complete installation instructions, advanced usage, API documentation, and examples, please visit: |
|
|
|
|
|
**[github.com/speridlabs/ORB](https://github.com/speridlabs/ORB)** |
|
|
|
|
|
## License |
|
|
|
|
|
Apache 2.0 - See [LICENSE](https://github.com/speridlabs/ORB/blob/main/LICENSE) |
|
|
|
|
|
## Acknowledgements |
|
|
|
|
|
Built upon the foundational work of the [DA-2](https://arxiv.org/abs/2509.26618). |
|
|
|
|
|
--- |
|
|
|
|
|
Made with ❤️ by [Sperid Labs](https://github.com/speridlabs) |
