Skip to main content
Daniel McCoy Stephenson
All projects

usd-synth-data

A procedural synthetic-dataset generator for computer vision: randomized OpenUSD scenes rendered headless with Blender Cycles into COCO-style images with exact boxes, masks and depth, fanned out as Kubernetes Jobs with measured throughput.

Python / OpenUSD / Kubernetes
Developer Tools
Active
GitHub

See it in 3D

A grid of three randomized tabletop scenes of coloured boxes, cans, balls, capsules and rings rendered by Strata, each shown beside its labelled bounding boxes, coloured instance masks and a rainbow depth map.

Three of twelve tabletop scenes from seed 2026, rendered by the Strata backend and laid out by scripts/contact_sheet.py: each image with its labelled boxes, instance masks and depth. Still drawn by usd-synth.