Unlocking AI Perception Gains with Iterative Synthetic Data Generation
In this webinar, Omar Maher and Phillip Thomas went over:
1. The importance of "playing with the knobs" – adjusting key parameters in synthetic datasets and observing their impact on AI perception model outcomes
2. A deep dive into critical dataset parameters, such as agent density/distribution, environment factors, sensor characteristics, and annotation specs
3. Real-world case studies demonstrating how iterative optimization has benefited various ML tasks, including 2D object detection, semantic segmentation, optical flow, and more
00:00: Introducing Omar and Phillip
01:12: How to Get Value Out of Synthetic Data
02:06: Agenda
02:54: Problems with Current ML Development Practices
04:23: Edge Case Detection using Synthetic Data
05:11: Parallel Domain’s Image Data (parking, trailers, debris, road signs, etc.)
05:45: Parallel Domain’s Sensors and Annotation types (labels)
06:19: Performance Improvements using Synthetic Data
07:45: Iterative Synthetic Dataset Design
09:15: Domain Gap Ontology
25:29 Parallel Domain SDK
28:46 Demo - using Synthetic Data for 2D Object Detection
32:11 Demo - Performance Iteration with Synthetic Data
35:11 Dataset Design - Optical Flow Example
39:54 Dataset Design - Semantic Segmentation UDA Example
43:43 Demo - Reducing Synthetic Data Iteration Times using Statistics
47:41 Why Parallel Domain
48:44 Parallel Domain at CVPR
51:24 Questions and Answers
If you would like to learn about Parallel Domain's platform, check out our website: https://paralleldomain.com/