US Patent 10977520 Training data collection for computer vision

Provided is a process that includes: determining that a training set lacks an image of an object with a given pose, context, or camera; composing, based on the determination, a video capture task; obtaining a candidate video; selecting a subset of frames of the candidate video as representative; determining that a given frame among the subset depicts the object from the given pose, context, or camera; and augmenting the training set with the given frame.