Spaces:
Running
Running
File size: 3,796 Bytes
dcc5ee3 7bd1cc6 dcc5ee3 7bd1cc6 dcc5ee3 7bd1cc6 dcc5ee3 7bd1cc6 dcc5ee3 7bd1cc6 b40a097 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 | ---
title: ICP Lab - 3D Point Cloud Alignment
emoji: 🎯
colorFrom: green
colorTo: gray
sdk: static
app_file: index.html
pinned: false
license: mit
tags:
- robotics
- computer-vision
- point-cloud
- education
short_description: Explore 3D point-cloud alignment step by step.
---
# ICP Lab: 3D point-cloud alignment
An interactive, CPU-friendly educational demonstration of rigid point-to-point Iterative Closest Point (ICP), inspired by [Dev Kumar's 3D Point Cloud Registration with ICP project](https://github.com/devkumar-projects/Classical-Computer-Vision-Toolkit/tree/main/3d-point-cloud-registration-icp). It generates a synthetic asymmetric surface, applies a known perturbation and lets visitors inspect the iterative alignment.
Drag to rotate the 3D view. Adjust rotation, translation and Gaussian noise; run the algorithm or advance one iteration at a time. The chart shows nearest-neighbour root mean square error (RMSE) in arbitrary units. The UI uses a perspective-projected view, while ICP itself fits a full 3D rigid transformation.
## Method and limits
Each iteration finds Euclidean nearest-neighbour correspondences by brute force, then estimates the least-squares rigid transformation via Horn's unit-quaternion formulation. This implementation is a new, independent educational adaptation, not an exact execution of the source notebook or Open3D. It does not benchmark hardware, reconstruct a real scene, or claim globally optimal alignment. ICP is local: initialization, point-cloud geometry and noise can produce incorrect correspondences and local minima. RMSE here is measured against the current nearest neighbours, not a guarantee of recovery of the true pose.
All points are generated in the browser; the demo requires no account, GPU, model downloads, or server-side inference. No data is uploaded. No libraries are needed.
## Run locally
Serve the directory with any static server; for example:
```bash
python3 -m http.server 8000
```
Then open `http://localhost:8000`. For an optional algorithm check, run `node test.mjs` (Node.js 18+).
## Publish as a Hugging Face Space
Create a **Static** Space and upload the contents of this folder at its root, preserving `README.md`, `index.html`, `style.css`, `app.js`, and `icp.js`. The YAML header configures the Space. No paid hardware is needed for a static Space. Inspect the finished public page before sharing.
## Further reading
- [Original project and notebook](https://github.com/devkumar-projects/Classical-Computer-Vision-Toolkit/tree/main/3d-point-cloud-registration-icp)
- [Related Kaggle notebook](https://www.kaggle.com/code/devkumarkaggle/3d-point-cloud-registration-with-icp)
- B. K. P. Horn, “Closed-form solution of absolute orientation using unit quaternions,” *Journal of the Optical Society of America A* 4(4), 629–642 (1987), DOI: [10.1364/JOSAA.4.000629](https://doi.org/10.1364/JOSAA.4.000629).
- P. J. Besl and N. D. McKay, “A method for registration of 3-D shapes,” *IEEE Transactions on Pattern Analysis and Machine Intelligence* 14(2), 239–256 (1992), DOI: [10.1109/34.121791](https://doi.org/10.1109/34.121791).
## License
The demo code is available under the MIT License. The referenced original project has its own [MIT license](https://github.com/devkumar-projects/Classical-Computer-Vision-Toolkit/blob/main/LICENSE).
## Related engineering work
Explore the [Engineering Workbench](https://huggingface.co/spaces/devkumarhf/Engineering-Workbench) for twelve source-linked projects, live RST synthesis and the original UR7e training-log explorer. The [Hugging Face collection](https://huggingface.co/collections/devkumarhf/robotics-perception-and-control-engineering-workbench-6afce1df9187d7a33c6da1ec) connects the interactive labs and documented data.
|