RandLA-Net Licencing Clarification #413
Description
We would like to use Open3D-ML within our organisation for semantic segmentation of point cloud data, but are concerned about the licencing for the RandLA-Net model.
The original RandLA-Net implementation is licenced under Creative Commons BY-NC-SA 4.0 International, which is a viral licence that also prohibits commercial use.
On the other hand, Open3D-ML is MIT-licenced, which would permit our use case. However, we have noted some locations in the Open3D-ML implementation of RandLA-Net that suggest that the Open3D-ML code is in fact derived from the original RandLA-Net, and therefore should also fall under the CC BY-NC-SA licence.
Examples (baseline commit b8e1eb1):
| Function | O3DML TensorFlow | O3DML PyTorch | QingyongHu (Original) | Notes |
|---|---|---|---|---|
random_sample |
Open3D-ML/randlanet.py#L312 | Open3D-ML/randlanet.py#L522 | RandLA-Net/RandLANet.py#L302 | Implementations are very similar, even to the code comments. |
nearest_interpolation |
Open3D-ML/randlanet.py#L334 | Open3D-ML/randlanet.py#L550 | RandLA-Net/RandLANet.py#L319 | As above, with the PyTorch implementation having more differences. |
gather_neighbour |
Open3D-ML/randlanet.py#L334 | Open3D-ML/randlanet.py#L361 | RandLA-Net/RandLANet.py#334 | The PyTorch implementation is significantly different to the original, whereas the TensorFlow implementation is virtually the same. |
| Core model | Open3D-ML/randlanet.py#518 | Open3D-ML/randlanet.py#L209 | RandLA-Net/RandLANet.py#L32 | Names of data items are the same without variation. |
While there do exist some areas of significant difference, the examples listed above give us pause regarding adoption of Open3D-ML due to RandLA-Net's licencing requirements. As such, we would appreciate some clarification or guidance from the project authors regarding Open3D-ML's licencing status.