A usable and extensible video annotation library for machine learning
License:
BSD-3 ClauseExternal links:
Video is captured in a wide variety of biology applications, including cell and animal behavior. Modern analysis of these data requires machine learning, used in problems such as cell or animal tracking and pose recognition, cell division detection, and behavior classification. As good training data is essential to the success of practical applications of machine learning, we need efficient methods for data annotation and visualization. We’ve been developing such applications for over about 15 years, and, for each new application we have worked on, we have had to write new, from-scratch interfaces. While each application has had differences, there are many commonalities that a well-designed library could exploit. Even more problematic is the fact that, throughout computer vision, we often artificially constrain ourselves to inference and learning over single time points because access to time series annotations is unavailable.
We have developed a general-purpose, extensible library for visualizing and annotating video data. It enables new machine learning applications to biology time series data, as it allows developers to efficiently collect annotations. One can quickly and easily incorporate this library into a new application, and tailor it to the details of the application, allowing for efficient labeling.