Some Questions Regarding Inference and Model

[dd-xx-dot]

I have been exploring your work on Sonata and Concerto, and I have a few questions regarding model inference and training details:

1、In PCA, I noticed that the inference model is set to mode=train. Could you please clarify why train is used instead of test, and what the purpose of test mode is?
https://github.com/facebookresearch/sonata/blob/18c09ff8d713494f78a8213792262b910977a65d/sonata/transform.py#L881C8-L881C47

2、You released a version that does not require color or normals https://huggingface.co/Pointcept/Sonata/blob/main/sonata_20251024.pth. How does its performance compare to the version that uses normals? Regarding sonata_20251024.pth, was this version trained on outdoor datasets? Would it be suitable for outdoor environments?

3、The sonata.pth released at https://huggingface.co/facebook/sonata/tree/main
, was it trained on any outdoor datasets?

4、Regarding the newly released Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations, was it trained on both indoor and outdoor data? Can it work using only point cloud data, without requiring color, normals, intensity, or other additional features?

5、Why were versions of Sonata and Concerto that do not use normals released, yet it is said to be a feature of Project Utonia? Is it because the performance of the no-normals versions in Sonata and Concerto is not very good?

Thank you very much for your time and assistance.

[apric0ts]

@dd-xx-dot Have you done any performance comparisons for the version that does not require color or normals w.r.t. outdoor datasets?

I would also appreciate an answer to 5

[Gofinge]

Sorry for missing this issue. Here is my response:

1. mode=test is a special multiple fragment ensemble testing strategy used in Pointcept. This is not necessary for simple mapping-back inference.
2. We didn't benchmark it, as this is an update from our ongoing research - Project Utonia. b But generally, I think it is pretty reliable, and can provide the best performance in case you don't have color / normal.
3. We have already moved the focus to the new project, so please check the Concerto demo, we have one weight trained with outdoor data.
4. Indoor and outdoor separately. But very soon, we will have Utonia trained on all of this data to shape one model.
5. It is a feature of Utonia, but we already pre-released this feature by recently updating the Concerto & Sonata weights. Please check Concerto demo 1, it contains a parameter to turn off normal / color.