Dataset Synthesis for Computer Vision with NVIDIA

A comprehensive pipeline for synthesizing high-quality computer vision datasets using Blender and nvisii to support advanced feature matching research with the LoFTR model.

Working in collaboration with NVIDIA’s Principal Research Manager Stan Birchfield and Dr. Jonathan Tremblay, this project developed a GitHub repository for synthesizing the YCB dataset, generating over 1,200 synthetic images with careful 3D object processing.

Key Achievements:

• Increased LoFTR model accuracy by 12% through careful synthetic data generation
• Improved model robustness by 10% using additional synthetic data
• Generated 1,200+ high-quality synthetic images
• Extracted vertex details for better 3D analysis
• Analyzed 150+ renders to elevate performance metrics by 15%

Technologies: Blender, nvisii, Python, Computer Vision, 3D Rendering, PyTorch

Mentors: Prof. Stan Birchfield (NVIDIA), Dr. Jonathan Tremblay (NVIDIA)

Affiliation: University of Washington & NVIDIA Research