What we have learned collecting and labelling data for robotics and computer vision: the rules, the edge cases and the trade-offs.
Egocentric data is first-person video of people doing real tasks. Here is why robotics teams use it, what makes raw footage usable, and how to judge an egocentric dataset.
Read the article →A practical walkthrough of turning first-person video into training data for vision-language-action models: the unit of work, the rule set, per-hand descriptions, review and the record schema.
Read the article →What first-person (egocentric) and third-person (exocentric) video each capture about human actions, where each falls short, and how to choose for a robot learning project.
Read the article →When to use bounding boxes, polygon segmentation, key points, classification, tagging and video annotation, with real project examples and how to keep quality consistent.
Read the article →