DataOps, MLOps, Computer Vision
summary
Argues that production computer vision fails or succeeds on data, not models, and defines three concrete data problems: curate high‑quality datapoints, label and audit at massive scale, and detect and adapt to data drift. Recommends a DataOps stack with acquisition, labeling, debugging (uses Great Expectations), augmentation, transformation, curation and edge‑case pipelines, orchestration with tools like apache airflow or Prefect, and claims Superb AI reduces dataset turnaround from months to days after building a labeling+audit platform.
AI-written summary · about
chapters (27)
- 0:00Introduction
- 1:21Outline
- 3:01MLOps
- 3:41Why DataOps
- 5:54DataOps Landscape
- 7:37Computer Vision
- 7:49Data is more important
- 8:29Data preparation challenges
- 9:27Building computer vision application altered
- 11:21Core principles of data ops
- 11:47Implement best practices for development
- 12:47Automate and orchestrate
- 13:47Test data quality
- 15:37Observability
- 17:49Data Metadata Model
- 19:56Empower Collaboration
- 22:40Data Office
- 23:16Data Pipeline
- 28:50Data Challenges
- 29:07Data Curation
- 30:40Label and Audit
- 33:22Data Drift
- 35:05Conclusion
- 35:15Modern Data Stack
- 37:00AI Infrastructure Alliance
- 38:27AI Infrastructure Stack
- 39:09Closing
related talks
01h 18mComputer Vision after the Victory of Data
04mUsing Computer Vision to Understand Big Visual Data
34mThe Recipe for Perfect Annotation in Computer Vision
23mComputer Vision Toolbox - Connected Computer Vision & AI Quality Inspection at Scale!
27mComputer vision for automated quality inspection
19mBuilding a Production Computer Vision Verification System
32mComputer Vision IRL: From Idea to On-Premise Deployment
01h 33mComputer Vision in Azure