ZiyuGuo99 / SAM2Point

The Most Faithful Implementation of Segment Anything (SAM) in 3D
https://sam2point.github.io/
Apache License 2.0
269 stars 14 forks source link

Full scene segmentation #2

Open dsvilarkovic opened 2 months ago

dsvilarkovic commented 2 months ago

Hey, really nice technical report, I will test some of the results on the GS3D pointclouds.

What's your opinion on using this approach to segment out the scene on large scale? I.e. have randomly sampled points and use them as prompts throughout the full pointcloud scene and get all the labelled objects? Do you think it will be too slow and do you have some better approaches for this?

ZiyuGuo99 commented 2 months ago

Thanks for your interest in our work! It's an important and practical application to segment an entire 3D scene.

We suggest initially dividing the 3D scene into multiple cubic partitions. Each partition can then be treated as an independent 3D input, using the cubic center as the point prompt, and segmented with SAM2Point. Finally, ensemble the results from all cubic partitions for a complete 3D scene segmentation.

This will allow for processing each cubic simultaneously, which enhances the efficiency and guarantees the granularity of segmentation.

We welcome further discussions on this topic!