ground truth of relative positions

[zebinchen018]

How to construct ground truth of relative positions between multiple 3D assets in a dataset?

[Mengmouxu]

During the dataset construction process, for multiple assets in the scene, we select one asset as the query asset. The position, rotation, and scale of the query asset are set to (0, 0, 0), no rotation, and 1, respectively. The position states of the remaining assets are represented as their relative positions, rotations, and sizes in the query asset's coordinate system. For detailed descriptions, refer to the relevant papers, and the code implementation logic can be referred to from the data loader during training.