superpoint boundary question

[kleinZGH]

✅ Code of conduct checklist

• 🌱 I am using the latest version of the code.
• 📙 I thoroughly went through the README, but could not find the feature I need there.
• 📘 I thoroughly went through the tutorial slides, notebook, and video, but could not find the feature I need there.
• 📗 I thoroughly went through the documentation, but could not find the feature I need there.
• 📜 I went through the docstrings and comments in the source code parts relevant to my problem, but could not find the feature I need there.
• 👩‍🔧 I searched for similar issues, but could not find the feature I need there.
• ⭐ Since I am showing interest in the project, I took the time to give the repo a ⭐ to show support. Please do, it means a lot to us !

🚀 The feature, motivation and pitch

Dear Author, thank you very much for your valuable contribution！
I am currently working on a point-cloud segmentation task based on SuperPoint Transformer, and I would like to focus specifically on boundaries. However, my understanding of superpoints is still rather shallow. Could you kindly advise whether boundary prediction in a superpoint-based framework is a topic worth exploring?

🔀 Alternatives

No response

📚 Additional context

No response

[drprojects]

Hello @kleinZGH if you are interested in the notion of boundaries in the scope of superpoints, you may want to read into superpoint/superpixel partition literature. In our work, we use cut-pursuit algorithm to produce a semantic oversegmentation of the point cloud. This is not the only partition method out there, other works have used VCCS, or recently a popular trend was—assuming you have 2D images with known pose wrt your point cloud—to project SAM masks onto 3D.

I you look into this, you can also have a look at the literature on 2D superpixels, which has been more studied than 3D superpoints. There are lots of papers out there but among the old ones you might want to know SLIC if you are not already familiar with it.

Finally, if your question is about segmenting boundaries for a specific task, I guess you could use the superpoint-based framework to do so efficiently. But our work is inherently designed for superpoint-wise predictions. In the specific case of this Superpoint Yransformer project, our model is designed and trained to classify the nodes of a graph. While for boundary-related tasks, you would probably want to classify edges of this graph, so that would require a bit of modifications.

Best
Damien