The Recipe for Perfect Annotation in Computer Vision
summary
Describes creating and releasing three large open datasets for computer vision—HaGRID, EasyPortrait, and Slovo—covering different CV tasks. Explains the annotation pipeline, quality-control practices, and tooling used to collect, label, and publish the datasets, and invites reports of errors via support@ontico.ru.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:49Annotation goals and projects
- 3:37Annotation methods in CV
- 6:28Crowdsourcing principles and overlap
- 7:53Annotation error analysis
- 8:58Framework pipeline overview
- 11:16Hard aggregation for boxes
- 11:57Hard aggregation for intervals
- 12:20Hard aggregation for masks
- 13:29Drop aggregation for errors
- 14:20Soft aggregation for improvement
- 17:13Confidence and threshold parameters
- 19:47Framework handling of annotations
- 22:26Parameter tuning for tasks
- 25:19Results and open source
- 26:26Q&A
related talks
51mHow to Build a Huge Dataset Without Spending a Fortune
01h 18mComputer Vision after the Victory of Data
29mLabeling Tools are Great, but What About Quality Checks?
18mCustomizable Computer Vision Expands Data Access Without Compromising Privacy
40mDataOps, MLOps, Computer Vision
27mDevelop Computer Vision Apps With Go
01h 08mMastering Visual AI with Vision-Language Models & Advanced Evaluation Techniques
30mIntroduction to Computer Vision with OpenCV