Hi, there. I’m one user who uses Omniverse for synthetic data generation.
When trying to create a synthetic dataset with Issac Sim,
I am wondering how multi-type labels can be accurately extracted, e.g., instance segmentation, semantic segmentation, 2D/3D bounding box, normal map, camera 3D pose, depth map, and so on.
What information is used in virtual simulations for automated annotation?
One of the things I’m expect to is that geometric information and materials in G-Buffer for Deferred Shading might be used for automatic annotation.
Because I discovered that clicking the Options button extracts accurate multi-type labels based on camera view port in an instant.
The reason I’m asking this is because I knew that early synthetic dataset generation paper[1-2] utilized G-Buffer, and I’m curious if Omniverse also adopts this method.
[1] https://arxiv.org/pdf/1608.02192
[2] https://openaccess.thecvf.com/content_ICCV_2017/papers/Richter_Playing_for_Benchmarks_ICCV_2017_paper.pdf