com.unity.perception

Why we included this project

Computer vision teams frequently hit a wall when they need large amounts of labeled training data but cannot collect or annotate enough real footage. This Unity package works around that by generating synthetic datasets from simulated scenes: you arrange labeled objects in the editor, capture RGB images together with ground truth such as bounding boxes, segmentation masks, and pose or keypoint labels, and use domain randomization to shift lighting, materials, and object placement between captures. Exports follow the SOLO schema and can be converted to COCO or pulled into tools like Voxel51 for inspection before training, so the output slots into an existing ML workflow. The catch is that Unity has discontinued the project and no longer supports it, meaning you would rely on community forks and GitHub issues rather than vendor help. For teams already invested in Unity it is still a capable reference implementation and a practical starting point for building a synthetic data pipeline.