Possible Inconsistency in Rotation Annotations (w2c) in Dataset: Objaverse_render_random

[AutonomierLong]

Hello, thank you for releasing this dataset.

I am currently trying to use the dataset you provided on Hugging Face:
https://huggingface.co/datasets/Viglong/Objaverse_render_random/tree/main

I noticed that the basic data format consists of a single rendered image (one object rendered on a black or white background) and a corresponding .npy file that stores the camera pose (world-to-camera, w2c).

However, after visualizing the rotation gizmo projected onto the image using the provided w2c matrices, I found that around 70% of the samples look correct, while the remaining ~30% seem to have incorrect pose annotations.

For example, the gizmo of this car below looks reasonable (x forward, y left, z up):

However, some other objects' gizmo looks odd:

I would like to ask:

• What could be the reason for this inconsistency?

• During training, did you further filter or clean this dataset, or apply any additional selection criteria on top of this released version?

Thank you very much for your time and help.