The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Simulation and Synthetic Data for Computer Vision with Batu Arisoy - TWiML Talk #281
Jul 9, 2019
In this engaging discussion, Batu Arisoy, a Research Manager at Siemens Corporate Technology, shares insights from his work on limited-data computer vision problems. He highlights innovative approaches to generating synthetic data, aiding object recognition for tasks like train maintenance. Batu also discusses collaborative workshops tackling class imbalances in computer vision, and a groundbreaking AI-user collaboration model with the Office of Naval Research that integrates natural language processing. Tune in for fascinating breakthroughs that blend AI, simulation, and user intent!
AI Snips
Chapters
Transcript
Episode notes
Spare Part Recognition
- Batu Arisoy's team developed a spare part recognition system for train maintenance.
- Service engineers use tablets to photograph parts and draw bounding boxes for automatic identification.
Synthetic Data and Sensor Modeling
- Synthetic data generation for computer vision should consider sensor characteristics.
- Modeling sensor noise, distortion, and other effects makes simulated data more realistic.
DepthSynth Pipeline
- DepthSynth models depth sensor behavior mathematically with modules for projector, camera, and object modeling.
- The pipeline also simulates post-processing steps like hole filling, generating realistic depth images.