Regarding point cloud feature extraction

[huassaue]

If I use the model to extract features from normalized point clouds of individual objects (rather than scenes), without color or normal information, will the model still work?

[Gofinge]

Good question. Our team is working on this. We will soon release a model called Concerto, but this still does not include normalized object point clouds. But for the next follow-up research of Concerto (codename Utonia), we will include object point cloud, indoor point cloud, and outdoor point cloud altogether to shape one pre-trained model. Please give us about half a year to make the model ready.

Concurrently, we also tried to directly apply the Sonata pre-training architecture to a single object dataset, and it should also work; we can get about 92% linear probing accuracy on ModelNet40.