Generated 3D synthetic data to train computer-vision networks
Research Engineer · New York, NY · 2015 – 2017
Golem, first cut (2016): “create body images from nothing (or, well, Blender and an .obj file)”. A headless Blender script rendered any body mesh, or a directory of them, into images. The renders were tagged synthetic in the data ledger and run through the images pipeline alongside the scraped photos, so the vision networks trained on both.
By December 2016 the renderer drove a SMPL body (per-vertex segmentation, shapes sampled from the model’s standard deviations, a spherical-harmonics lighting shader) through CMU motion-capture clips and composited it onto background photos, frame by frame.
Golem grew into the synthetic-humans package that Amazon Go’s data framework later built on.
Python · Blender