Building Training Data for Autonomous Driving
summary
Company pitches training data as a service and reports $27M raised while explaining concrete tooling choices: initial bounding-box and full-image segmentation workflows fail for driving data, so tasks break into microtasks and recursion to boost recall. A pipeline aggregates many per-object masks, scores contributors with a user reputation model, then applies CV pre-segmentation into superpixels so annotators paint labels, producing pixel-accurate semantic masks at scale and lower cost.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 6:27Customers
- 6:49Public Data
- 8:09Sample Images
- 8:57Annotation Tools
- 10:25MicroTasks
- 11:29Recursion
- 12:35PreSegmenting
- 14:07Recommendations
- 14:52Trust
related talks
41mAI applications in autonomous driving
26mAI for Autonomous mobile robots
22mUrban Artificial Intelligence in Amsterdam
23mBuilding high-quality computer vision models using only a few examples
36mDeep Learning - Segmentation, Identification and Classification
18mCustomizable Computer Vision Expands Data Access Without Compromising Privacy
58mUsing Unity to Generate Synthetic data and Accelerate Computer Vision Training
45mTech Tutorial: Image segmentation