Autorevise

What is it about?

This work addresses the issue of video annotation quality for activity recognition applications. It proposes a method to detect errors and correct the starting and ending locations (localization) of an annotated activity. It proposes two methods to assess the quality of activity annotations provided by humans.

Featured Image

Why is it important?

When humans annotated a large set of video data, they are prone to errors. Precise localization (identifying the correct starting and ending frames in a video) is an especially difficult task for a human to perform. Other machine learning researchers have reported evidence of incorrect annotations impacting the performance of machine learning algorithms. Having methods to assess the overall quality of class labels and being able to correct temporal boundaries can improve the robustness of machine learning models.