bpycv

Why we included this project

Getting the right training data for vision models is often the bottleneck. For segmentation, depth, or 6DoF pose work, real images can be scarce and labeling them expensive. bpycv treats Blender as a synthetic-data generator: a short Python script renders an RGB scene with semantic, instance, and panoptic masks, a depth map, and 6DoF pose ground truth in one call. Domain randomization is built in, pulling random HDRI backgrounds from HDRI Haven, ShapeNet models as distractors, and textures from Texture Haven, so generated scenes don't end up looking identical. It works through Blender's native Python API and supports Docker, which keeps it approachable for researchers and small teams already at home in Blender, and output converts to Cityscapes format to cut a common integration step before training.