Object Coordinate Rendering

Renders synthetic RGB-D training data from BOP-format object pose datasets by projecting 3D models with ground-truth poses.

Input

No data selection is required. Configure paths to a BOP training split and its object models on disk.

Output

Depends on Export location:

  • Suite — image sets added to the data model: rendered object coordinates, optional surface depth, and optional segmentation masks.

  • Disk — HDF5 files written under subfolders of the BOP data directory (rendered_object_coordinates, rendered_depth, rendered_mask per sequence).

Description

The algorithm walks numeric sequence folders under the BOP data path, reads scene_gt.json and scene_camera.json per sequence, loads meshes from the models directory, and renders each annotated object instance with SimulatorRenderer.

Typical layout:

  • BOP data directory — path to a split such as …/ycbv/train_real containing sequence subfolders and scene JSON files.

  • BOP models directory — path to the corresponding models folder with one mesh file per object ID.

Render segmentation mask — output a binary mask image per instance.

Render surface depth — output a depth map per instance.

Fix K of PBR — apply a fixed −0.5 px principal-point offset to match the original BOP PBR renders. Use only with the official BOP dataset; leave disabled otherwise.

Export location — Suite loads results into the application (avoid for large batches). Disk writes HDF5 files next to the dataset for offline training pipelines.